Productive AI starts with the foundation of your business.
Processes, data, systems and compliance decide whether AI lands in day-to-day work. We assess these four pillars with you before the choice of tools comes up, and take you from there to production use with a holistic framework built around RAG and the training of your people.
You decide how we work together: consulting that enables your team, or a service where we take on the task for you. The goal stays the same: AI that becomes productive in your business.
Area 01
AI consulting for SMEs
From concept to production — with one holistic framework.
AI audit — where you stand along the four pillars, the biggest value potential and compliance in view. From CHF 800 excl. VAT when combined with training, including modules 1 and 2 of the workplace AI training, free of charge for two people. See where AI pays off
AI strategy & concept for your company
AI assistant — a personal assistant for management and specialist roles
Integration with your systems — AI where the work happens: Microsoft 365, Google Workspace, Slack and your business applications
Corporate LLM — local model, your data in your own house
Raise quality with AI
Grow revenue — AI in sales: sharpen customer outreach, match enquiries automatically, produce tailored offers largely hands-free
Holistic AI framework with RAG at the centre
Prompt/context engineering & data management
Software development with AI: frontier models & agentic workflows
Training & enablement — workplace AI training with an AI tutor. Go to course
Tools only work as well as the foundation they stand on. The framework describes the technology — the foundation describes your business: processes, data, systems, compliance.
Processes & organisation
Only documented processes can be delegated — to people and to AI alike. What lives only in people's heads, no tool can take over.
Data hygiene
AI is only as good as the data it can reach. Scattered, outdated repositories are the quietest AI blocker — no tool fixes them.
Infrastructure
It's not a system's age that decides AI readiness, but data access: export or interface available — or an island.
Compliance
AI and data protection get along — but not by themselves. Clear ground rules for what may go into an AI tool and what never should — with the Swiss revDSG in view.
An honest assessment along these four pillars is the first step of every AI audit — or work it out yourself in “The Groundwork”, module 2 of our workplace AI training.
Holistic AI framework
RAG at the centre — everything else aligns to it
A coherent approach, not point tools. Your knowledge, your data and the right models interlock.
RAG
Retrieval-augmented generation on your own knowledge — the architectural core.
Data management
Where your data lives and how it's governed — in Switzerland, the EU or worldwide, secure and auditable.
Context engineering
What the model sees: prompts, context and guardrails — designed on purpose.
Frontier vs. local models
Hosted frontier models and local models — combined by cost, speed, capability and sovereignty.
RAG inserts one step between question and answer: first search your curated knowledge base, then let the model answer from it. A generic model becomes company-specific knowledge.
R
Retrieve
Search
The question is matched against the knowledge base — only reviewed content is found.
A
Augmented
Enrich
The evidence goes into the prompt: “Here’s context, answer with it.”
G
Generation
Answer
Answer based on the sources. Good source, good answer. Hence: curated.
The “Intelligence” in Artificial Intelligence is, at its core, the structured processing of data and information — closer to “knowledge” or an “information edge” than to universal human reason.
An immense knowledge network
AI recognises patterns and links connections with great precision — provided we give it the right context and specific information.
The human decides
AI delivers the groundwork. The final logical check, critical thinking and strategic judgement remain irreplaceable.
Use the technology — but keep your own mind switched on.
Why iConference
Consulting that reaches production
SME focus
Pragmatic, budget-aware and aimed at measurable value.
Data sovereignty
Sensitive data stays in Switzerland — frontier models are used where best-in-class AI performance is required, safeguarded by anonymisation and clear guardrails.
Concept to production
Not just strategy — we get AI into productive use.
We bring AI to where it should work: into the daily life of your company — from concept all the way to production.
— ICONFERENCE AG · ZUG
Origin
We offer Swiss made AI
iConference AG carries the labels «swiss made software +AI» and «swiss digital services +AI». We develop our AI solutions in Zug and run them in Swiss data centres, on international platforms or on the client’s own premises, depending on the use case.
Our work ranges from consulting and individual AI services through to our own software: the Sovereign Cockpit, which anonymises confidential documents before they leave the building, and the corporate AI training course with its AI tutor. We work with open models on Swiss inference (Apertus, Gemma, GLM, Qwen), with frontier models from the large providers, with RAG and agentic workflows, and on request with a local model on the client’s own hardware.
An independent entrepreneur since 1997, Georges Leuenberger knows first-hand how lonely leadership can be. How do I win this deal? Where do I find new customers? How do I optimise operations, find and realise cost savings, and deliver better than the competition? That's exactly where iConference comes in.
Career
Business economist HF
10 years of software development on IBM large-scale systems, incl. business software for the juvenile courts of the Canton of Bern and the payout system of the Swiss unemployment funds
Lotus Notes consulting
Managing Director Switzerland of PictureTel — then the world leader in high-end videoconferencing
Own company: system integrator & managed service provider for telepresence environments
Consulting in Unified Communications and “The Smart Way to Work”
Interim COO for the Mittelstand
AI evangelist for sustainable use — measurable productivity with better quality
What you hire us for
You hire us for three things: a foundation that holds (processes, data, systems, compliance); technology such as RAG that works on your own knowledge; and the training of your people. We consider the training the most important part: the best technology achieves little if people do not understand it and use it in their daily work. Depending on the mandate, we bring in specialised partners. We disclose where AI runs in our own business.
AI NewsCurated and selected by Claude Opus · always current
AI news, always current
We gather the developments that matter from the AI world — short, contextualised, no noise.
Tools14 Nov 2026
Canonical presents the Inference Manager for local AI models
According to heise online, Canonical used the Ubuntu Summit to present the Inference Manager, a new Ubuntu tool for managing locally installed AI models. For SMEs: anyone running AI models locally on Ubuntu machines should follow whether the tool makes managing those models easier.
Replit agent buys supplies and runs Google Ads, only after approval
According to AI Insiders, Replit's coding agent can now set up Google Ads campaigns and shop through Stripe's payment service Link with Visa, with shopping starting in the US and covering one-time purchases only. Nothing is bought or launched until the owner approves, Replit says. For SMEs: anyone giving an agent access to payment methods or the ads account should first decide who approves which spending.
Gartner figures: AI drives IT spending to new highs in 2027
According to Netzwoche, worldwide IT spending excluding telecom services will approach the 6 trillion US dollar mark in 2027, based on Gartner figures. Netzwoche names data centres and AI infrastructure as the main drivers, while rising memory prices are causing tension in the hardware market. For SMEs: anyone planning to replace servers or computers in 2027 should factor the rising memory prices into the budget early.
According to OpenAI, agents in ChatGPT Work, Codex and dots can complete purchases with ChatGPT Wallet, each with the user's approval and with existing card protections. For SMEs: anyone letting agents buy on the company's behalf should first decide who approves purchases and which card is stored for them.
United Airlines plans to restrict bookings by third-party AI agents
According to The Daily Agentic, United Airlines has announced that it will restrict unauthorised third-party AI agents from accessing customer accounts and booking flights, including through travel agencies and booking platforms. United cites exposed credentials, unintended purchases and check-in declarations that a human must confirm. For SMEs: anyone having AI agents book business travel should first check whether the airline allows such agents.
Gemini 4 Hydrogen: Google's lowest-cost model in the Gemini 4 family
According to Simon Willison, Google has launched Gemini 4 Hydrogen, the lowest-cost Gemini 4 model, which according to heise powers AI Mode in Google Search. Willison lists Hydrogen at 1.50 dollars per million input tokens and 7.50 dollars per million output tokens, and the larger Carbon and Argon at 2 and 10 dollars until 30 November and 4 and 20 dollars from 1 December. For SMEs: anyone using Argon via the API should check before 1 December whether the cheaper Hydrogen is enough for their tasks.
Grok Bot for Mac gains integration with Apple Mail and Messages
According to 9to5Mac, Grok Bot for Mac gains integration with Apple Mail, Messages, Notes and more, and according to BigGo Finance Grok Bot has also reached version 1.0. So far only the headlines of the two outlets are known; the headlines do not say what rights the bot gets in Mail and Messages. For SMEs: anyone allowing AI agents on company computers should first decide which mailboxes and chats such an agent may access.
Canton of Zug funds two AI security professorships at ETH Zurich
According to Netzwoche, ETH Zurich is getting two new professorships that bring together artificial intelligence and cybersecurity. The canton of Zug is investing up to 55 million francs over 25 years and, from 2027, will also work with ETH's Zurich Information Security and Privacy Center. For SMEs: anyone introducing AI should plan how to secure the systems from the start.
CrowdStrike unveils a kill switch for rogue AI agents
According to heise online, CrowdStrike presented a kill switch for its Falcon Guardian platform at its Fal.Con conference. According to heise, it is meant to stop AI agents that have gone rogue. For SMEs: anyone giving AI agents access to systems should decide in advance how an agent can be stopped immediately and who may trigger that.
Grok Build: flaw exposed data from thousands of apps
According to Technology Org, a flaw in Grok Build, the app-building tool from Elon Musk's company SpaceXSI, exposed data from thousands of apps built with it, including apps of SpaceXSI itself. According to a count by Claroc cited by AI Insiders, 35,578 apps built with Grok can be opened without a login; so far only the Technology Org headline and a short item in AI Insiders are known. For SMEs: anyone building apps with an AI tool should check before going live whether they can be reached without signing in.
Gemini Decision 1: Google's model for yes/no and choice decisions
According to Simon Willison, Google offers Gemini Decision 1 through the Gemini Decisions API, a model that answers yes/no questions, multiple-choice questions and rubric-based scoring. According to Willison, input costs 10 cents per million tokens, output is free, and the model handles images and audio as well as text. For SMEs: anyone automating recurring decisions, such as sorting incoming emails, should first measure the hit rate on their own cases.
CERT-EU warns of rapidly rising attacks using agent frameworks
According to heise online, CERT-EU, the cybersecurity service for the EU institutions, warns of a rapid increase in attacks using agent frameworks such as Artex. According to heise, autonomous AI agents carry out complex attacks largely on their own, and experts warn of a new wave of automated attacks. For SMEs: the warning is a reason to review updates, access rights and backups now.
Compiled automatically · always up to date · summaries may be translated — the linked source can be in another language.
Frequently asked questions about AI for mid-sized companies
What does iConference AG do?
We are an AI consultancy based in Zug, working with SMEs and mid-sized companies in Switzerland and Europe. We start with the foundation of your business: processes, data, systems and compliance. The technology builds on that, with RAG on your own knowledge. Two areas: consulting from assessment to production use, and AI services, where we handle individual tasks for you.
What are the four pillars of an AI deployment?
Processes and organisation, data hygiene, infrastructure and compliance. Tools only work as well as the foundation they stand on. A process that lives only in people's heads cannot be delegated, neither to people nor to AI. Scattered and outdated file storage is the quietest obstacle to AI. An honest assessment along these four pillars is the first step.
Where should a company start with AI?
With the foundation. The framework describes the technology, the foundation describes your business. Only once processes are documented, data is findable and the ground rules are settled will an AI tool carry weight. So the assessment comes first, then the choice of a use case with visible benefit.
What is an AI audit?
A structured assessment along four pillars: processes and organisation, data hygiene, infrastructure and compliance. We check which processes are documented and therefore delegable, where your data sits, whether your systems offer an export or an interface, and what ground rules apply to using AI. The result shows the largest opportunities and the open points.
What does an AI audit cost?
An AI audit combined with training starts at CHF 800, excl. VAT. This includes modules 1 and 2 of our workplace AI training, free of charge for two people. The audit delivers the assessment along the four pillars, the training brings your team to a shared level. A concrete quote follows after a short conversation.
Does our data stay in Switzerland?
Sensitive data stays in Switzerland. We use frontier models where their capability makes the difference, backed by anonymisation and clear guardrails. For data storage you choose between hosting in Switzerland, in the EU or worldwide. Which option fits depends on how protection-worthy your data is.
Which companies does iConference work with?
With SMEs and mid-sized companies in Switzerland and Europe. The approach is pragmatic and budget-aware: we work with what already exists in the business and aim AI use at measurable benefit. The company is based in Zug. We work on site in the canton of Zug, the Zurich area, Central Switzerland and the Aarau area, and by video elsewhere in Switzerland and in Europe.
What is the difference between AI consulting and AI services?
In consulting we enable your business to use AI productively on its own: audit, strategy, framework, training. In services we take on the task ourselves, for example revising contracts, restructuring product catalogues, turning documents into structured data or answering complex RFPs. For that you need no AI environment of your own.
Let's go. Bring AI into your business.
Tell us where you are today — we'll show you the next concrete step.
0,000 and is due to ship this year, but for most tasks it still needs a remote human operator; Unitree sells a humanoid for under $6,000; Elon Musk holds out the Optimus for the end of 2027 at about 0,000, after scaling back his goal of thousands of robots in Tesla factories by the end of 2025 to "some" doing simple tasks; an attempt by Google DeepMind to have a Gemini Robotics model pack the ingredients for a mushroom risotto into a basket failed; as reasons the article names a lack of training data from the physical world, the variety of real environments, unsolved navigation in new spaces and the requirement that a practical robot must almost always succeed, while many demonstrations are remote-controlled or scripted; Morgan Stanley nevertheless expects nearly one billion human-like robots by 2050 and a market above $5 trillion; for SMEs this means: for logistics with clearly defined grips, first trials are running in large companies, while for workshops, care or households no deployable solution is to be expected in the next few years; anyone reviewing offers should ask for the success rate without remote control.
Source: MIT Technology Review
Models7 Oct 2026
Falcon ASR: speech recognition with 1.6 billion parameters for Arabic, English, French, Spanish and Portuguese, 5.74 percent word error rate in English
According to the blog of the Technology Innovation Institute in Abu Dhabi on Hugging Face on 7 October, Falcon ASR has been released, a speech recognition model with 1.6 billion parameters built on Falcon3-Audio; it handles Arabic with a focus on the Emirati dialect as well as English, French, Spanish and Portuguese with the same weights and no language flag, and German is not included; under the protocol of the Open Universal Arabic ASR Leaderboard it averages 20.92 percent word error rate across six test sets against 23.17 percent for the best published result; in an internal evaluation for Emirati it scores 22.73 percent word error rate and 10.19 percent character error rate, 4.07 points better than Qwen3-Omni; on seven public English test sets of the Open ASR Leaderboard it scores 5.74 percent word error rate; it provides word-level timestamps, and training included background noise, overlapping speech, music, reverberation, telephony effects and changes in speed and pitch; it can be tried in a demo space with clips of up to 60 seconds, API access and applications are announced, and the post names no licence and no hardware requirements; for SMEs this means: for French-language calls, meetings and phone recordings another model is in the running, for German and Swiss German it brings nothing; the figures come from the vendor, and without a licence a production deployment cannot be planned yet.
Claude Haiku 5.5: $0.10 per million input tokens up to 100,000 tokens, Sonnet 5.5 cache reads halved
According to Anthropic on 7 October, Claude Haiku 5.5 has been released, the new version of its smallest and cheapest model class; for prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, above that $0.50 and .50, while Haiku 4.5 was at and $5; Anthropic puts the saving at around 75 percent on average, heise online at up to 90 percent for short prompts; it is the first Haiku model with an adjustable effort setting from Low to Max, and the Python and TypeScript SDKs gain computer use and browser use in beta; it runs on the Claude Platform, AWS, Google Cloud and Microsoft Azure; on OSWorld 2.1 (offline subset) Anthropic reports 72.4 percent, on Humanity's Last Exam 45.9 percent without and 57.4 percent with tools; heise online reports a context window of one million tokens instead of 200,000 and up to 128,000 output tokens, which Anthropic's page does not state; the new tokenizer uses slightly more tokens per task according to Anthropic, heise puts it at around 30 percent more, Simon Willison measures around 25 percent more on a long prompt; according to heise, citing Artificial Analysis, Haiku 5.5 scores 43 points on the Intelligence Index, GPT-6 Luna 38 and Sonnet 5.5 56; on the same day Anthropic halved the cache read price for Sonnet 5.5 from $0.20 to $0.10 per million tokens, which is meant to make agentic tasks around 20 percent cheaper; Max and Team subscribers receive a monthly API credit of 00, 00 or up to $500 this week; for SMEs this means: classification, summaries, database queries and pre-sorting at volume become considerably cheaper; the extra consumption of the new tokenizer and the price step above 100,000 tokens belong in your own calculation, and the test figures come from the vendor.
EmbeddingGemma 2: Google releases a multimodal embedding model for on-device use under Apache 2.0
According to heise online on 7 October, Google has released EmbeddingGemma 2, an embedding model that maps text, code, images, audio and video into a shared vector space, so that video recordings can be searched with a spoken query, for example; the model builds on the Gemma 4 architecture and is modular: 270 million parameters for text and code, 170 million for the image encoder, 300 million for the audio encoder, 740 million when fully loaded; vectors have 768 dimensions by default and can be cut to 512, 256 or 128, and the context window is 8,192 tokens, four times that of the predecessor; according to Google, one input can hold about 5.5 minutes of audio, 29 images or 58 video frames; on a Pixel 11 Pro the quantised model needs about 191 MB of active RAM for the text weights and about 567 MB for the full multimodal variant, by Google's account; on the MTEB Code benchmark the score rises from 68.76 to 78.68 points, again by Google's figures; the weights are on Hugging Face and Kaggle under Apache 2.0, and Google names LiteRT, MediaPipe, Transformers and llama.cpp as integrations; Simon Willison noted on 6 October that open weights matter especially for embedding models, because stored vectors have to be recomputed when the model changes; the article does not name supported languages; for SMEs this means: anyone building a local search across documents, images or recordings gets a small model with a free licence that runs on the device and keeps data in house; the scores come from Google, so quality on your own data should be checked before deployment.
Compiled automatically · always up to date · summaries may be translated — the linked source can be in another language.
Frequently asked questions about AI for mid-sized companies
What does iConference AG do?
We are an AI consultancy based in Zug, working with SMEs and mid-sized companies in Switzerland and Europe. We start with the foundation of your business: processes, data, systems and compliance. The technology builds on that, with RAG on your own knowledge. Two areas: consulting from assessment to production use, and AI services, where we handle individual tasks for you.
What are the four pillars of an AI deployment?
Processes and organisation, data hygiene, infrastructure and compliance. Tools only work as well as the foundation they stand on. A process that lives only in people's heads cannot be delegated, neither to people nor to AI. Scattered and outdated file storage is the quietest obstacle to AI. An honest assessment along these four pillars is the first step.
Where should a company start with AI?
With the foundation. The framework describes the technology, the foundation describes your business. Only once processes are documented, data is findable and the ground rules are settled will an AI tool carry weight. So the assessment comes first, then the choice of a use case with visible benefit.
What is an AI audit?
A structured assessment along four pillars: processes and organisation, data hygiene, infrastructure and compliance. We check which processes are documented and therefore delegable, where your data sits, whether your systems offer an export or an interface, and what ground rules apply to using AI. The result shows the largest opportunities and the open points.
What does an AI audit cost?
An AI audit combined with training starts at CHF 800, excl. VAT. This includes modules 1 and 2 of our workplace AI training, free of charge for two people. The audit delivers the assessment along the four pillars, the training brings your team to a shared level. A concrete quote follows after a short conversation.
Does our data stay in Switzerland?
Sensitive data stays in Switzerland. We use frontier models where their capability makes the difference, backed by anonymisation and clear guardrails. For data storage you choose between hosting in Switzerland, in the EU or worldwide. Which option fits depends on how protection-worthy your data is.
Which companies does iConference work with?
With SMEs and mid-sized companies in Switzerland and Europe. The approach is pragmatic and budget-aware: we work with what already exists in the business and aim AI use at measurable benefit. The company is based in Zug. We work on site in the canton of Zug, the Zurich area, Central Switzerland and the Aarau area, and by video elsewhere in Switzerland and in Europe.
What is the difference between AI consulting and AI services?
In consulting we enable your business to use AI productively on its own: audit, strategy, framework, training. In services we take on the task ourselves, for example revising contracts, restructuring product catalogues, turning documents into structured data or answering complex RFPs. For that you need no AI environment of your own.
Let's go. Bring AI into your business.
Tell us where you are today — we'll show you the next concrete step.
0,000 and is due to ship this year, but for most tasks it still needs a remote human operator; Unitree sells a humanoid for under $6,000; Elon Musk holds out the Optimus for the end of 2027 at about 0,000, after scaling back his goal of thousands of robots in Tesla factories by the end of 2025 to "some" doing simple tasks; an attempt by Google DeepMind to have a Gemini Robotics model pack the ingredients for a mushroom risotto into a basket failed; as reasons the article names a lack of training data from the physical world, the variety of real environments, unsolved navigation in new spaces and the requirement that a practical robot must almost always succeed, while many demonstrations are remote-controlled or scripted; Morgan Stanley nevertheless expects nearly one billion human-like robots by 2050 and a market above $5 trillion; for SMEs this means: for logistics with clearly defined grips, first trials are running in large companies, while for workshops, care or households no deployable solution is to be expected in the next few years; anyone reviewing offers should ask for the success rate without remote control.
Source: MIT Technology Review
Models7 Oct 2026
Falcon ASR: speech recognition with 1.6 billion parameters for Arabic, English, French, Spanish and Portuguese, 5.74 percent word error rate in English
According to the blog of the Technology Innovation Institute in Abu Dhabi on Hugging Face on 7 October, Falcon ASR has been released, a speech recognition model with 1.6 billion parameters built on Falcon3-Audio; it handles Arabic with a focus on the Emirati dialect as well as English, French, Spanish and Portuguese with the same weights and no language flag, and German is not included; under the protocol of the Open Universal Arabic ASR Leaderboard it averages 20.92 percent word error rate across six test sets against 23.17 percent for the best published result; in an internal evaluation for Emirati it scores 22.73 percent word error rate and 10.19 percent character error rate, 4.07 points better than Qwen3-Omni; on seven public English test sets of the Open ASR Leaderboard it scores 5.74 percent word error rate; it provides word-level timestamps, and training included background noise, overlapping speech, music, reverberation, telephony effects and changes in speed and pitch; it can be tried in a demo space with clips of up to 60 seconds, API access and applications are announced, and the post names no licence and no hardware requirements; for SMEs this means: for French-language calls, meetings and phone recordings another model is in the running, for German and Swiss German it brings nothing; the figures come from the vendor, and without a licence a production deployment cannot be planned yet.
Claude Haiku 5.5: $0.10 per million input tokens up to 100,000 tokens, Sonnet 5.5 cache reads halved
According to Anthropic on 7 October, Claude Haiku 5.5 has been released, the new version of its smallest and cheapest model class; for prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, above that $0.50 and .50, while Haiku 4.5 was at and $5; Anthropic puts the saving at around 75 percent on average, heise online at up to 90 percent for short prompts; it is the first Haiku model with an adjustable effort setting from Low to Max, and the Python and TypeScript SDKs gain computer use and browser use in beta; it runs on the Claude Platform, AWS, Google Cloud and Microsoft Azure; on OSWorld 2.1 (offline subset) Anthropic reports 72.4 percent, on Humanity's Last Exam 45.9 percent without and 57.4 percent with tools; heise online reports a context window of one million tokens instead of 200,000 and up to 128,000 output tokens, which Anthropic's page does not state; the new tokenizer uses slightly more tokens per task according to Anthropic, heise puts it at around 30 percent more, Simon Willison measures around 25 percent more on a long prompt; according to heise, citing Artificial Analysis, Haiku 5.5 scores 43 points on the Intelligence Index, GPT-6 Luna 38 and Sonnet 5.5 56; on the same day Anthropic halved the cache read price for Sonnet 5.5 from $0.20 to $0.10 per million tokens, which is meant to make agentic tasks around 20 percent cheaper; Max and Team subscribers receive a monthly API credit of 00, 00 or up to $500 this week; for SMEs this means: classification, summaries, database queries and pre-sorting at volume become considerably cheaper; the extra consumption of the new tokenizer and the price step above 100,000 tokens belong in your own calculation, and the test figures come from the vendor.
EmbeddingGemma 2: Google releases a multimodal embedding model for on-device use under Apache 2.0
According to heise online on 7 October, Google has released EmbeddingGemma 2, an embedding model that maps text, code, images, audio and video into a shared vector space, so that video recordings can be searched with a spoken query, for example; the model builds on the Gemma 4 architecture and is modular: 270 million parameters for text and code, 170 million for the image encoder, 300 million for the audio encoder, 740 million when fully loaded; vectors have 768 dimensions by default and can be cut to 512, 256 or 128, and the context window is 8,192 tokens, four times that of the predecessor; according to Google, one input can hold about 5.5 minutes of audio, 29 images or 58 video frames; on a Pixel 11 Pro the quantised model needs about 191 MB of active RAM for the text weights and about 567 MB for the full multimodal variant, by Google's account; on the MTEB Code benchmark the score rises from 68.76 to 78.68 points, again by Google's figures; the weights are on Hugging Face and Kaggle under Apache 2.0, and Google names LiteRT, MediaPipe, Transformers and llama.cpp as integrations; Simon Willison noted on 6 October that open weights matter especially for embedding models, because stored vectors have to be recomputed when the model changes; the article does not name supported languages; for SMEs this means: anyone building a local search across documents, images or recordings gets a small model with a free licence that runs on the device and keeps data in house; the scores come from Google, so quality on your own data should be checked before deployment.
Compiled automatically · always up to date · summaries may be translated — the linked source can be in another language.
Frequently asked questions about AI for mid-sized companies
What does iConference AG do?
We are an AI consultancy based in Zug, working with SMEs and mid-sized companies in Switzerland and Europe. We start with the foundation of your business: processes, data, systems and compliance. The technology builds on that, with RAG on your own knowledge. Two areas: consulting from assessment to production use, and AI services, where we handle individual tasks for you.
What are the four pillars of an AI deployment?
Processes and organisation, data hygiene, infrastructure and compliance. Tools only work as well as the foundation they stand on. A process that lives only in people's heads cannot be delegated, neither to people nor to AI. Scattered and outdated file storage is the quietest obstacle to AI. An honest assessment along these four pillars is the first step.
Where should a company start with AI?
With the foundation. The framework describes the technology, the foundation describes your business. Only once processes are documented, data is findable and the ground rules are settled will an AI tool carry weight. So the assessment comes first, then the choice of a use case with visible benefit.
What is an AI audit?
A structured assessment along four pillars: processes and organisation, data hygiene, infrastructure and compliance. We check which processes are documented and therefore delegable, where your data sits, whether your systems offer an export or an interface, and what ground rules apply to using AI. The result shows the largest opportunities and the open points.
What does an AI audit cost?
An AI audit combined with training starts at CHF 800, excl. VAT. This includes modules 1 and 2 of our workplace AI training, free of charge for two people. The audit delivers the assessment along the four pillars, the training brings your team to a shared level. A concrete quote follows after a short conversation.
Does our data stay in Switzerland?
Sensitive data stays in Switzerland. We use frontier models where their capability makes the difference, backed by anonymisation and clear guardrails. For data storage you choose between hosting in Switzerland, in the EU or worldwide. Which option fits depends on how protection-worthy your data is.
Which companies does iConference work with?
With SMEs and mid-sized companies in Switzerland and Europe. The approach is pragmatic and budget-aware: we work with what already exists in the business and aim AI use at measurable benefit. The company is based in Zug. We work on site in the canton of Zug, the Zurich area, Central Switzerland and the Aarau area, and by video elsewhere in Switzerland and in Europe.
What is the difference between AI consulting and AI services?
In consulting we enable your business to use AI productively on its own: audit, strategy, framework, training. In services we take on the task ourselves, for example revising contracts, restructuring product catalogues, turning documents into structured data or answering complex RFPs. For that you need no AI environment of your own.
Let's go. Bring AI into your business.
Tell us where you are today — we'll show you the next concrete step.
0,000 and is due to ship this year, but for most tasks it still needs a remote human operator; Unitree sells a humanoid for under $6,000; Elon Musk holds out the Optimus for the end of 2027 at about 0,000, after scaling back his goal of thousands of robots in Tesla factories by the end of 2025 to "some" doing simple tasks; an attempt by Google DeepMind to have a Gemini Robotics model pack the ingredients for a mushroom risotto into a basket failed; as reasons the article names a lack of training data from the physical world, the variety of real environments, unsolved navigation in new spaces and the requirement that a practical robot must almost always succeed, while many demonstrations are remote-controlled or scripted; Morgan Stanley nevertheless expects nearly one billion human-like robots by 2050 and a market above $5 trillion; for SMEs this means: for logistics with clearly defined grips, first trials are running in large companies, while for workshops, care or households no deployable solution is to be expected in the next few years; anyone reviewing offers should ask for the success rate without remote control.
Source: MIT Technology Review
Models7 Oct 2026
Claude Haiku 5.5: $0.10 per million input tokens up to 100,000 tokens, Sonnet 5.5 cache reads halved
According to Anthropic on 7 October, Claude Haiku 5.5 has been released, the new version of its smallest and cheapest model class; for prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, above that $0.50 and .50, while Haiku 4.5 was at and $5; Anthropic puts the saving at around 75 percent on average, heise online at up to 90 percent for short prompts; it is the first Haiku model with an adjustable effort setting from Low to Max, and the Python and TypeScript SDKs gain computer use and browser use in beta; it runs on the Claude Platform, AWS, Google Cloud and Microsoft Azure; on OSWorld 2.1 (offline subset) Anthropic reports 72.4 percent, on Humanity's Last Exam 45.9 percent without and 57.4 percent with tools; heise online reports a context window of one million tokens instead of 200,000 and up to 128,000 output tokens, which Anthropic's page does not state; the new tokenizer uses slightly more tokens per task according to Anthropic, heise puts it at around 30 percent more, Simon Willison measures around 25 percent more on a long prompt; according to heise, citing Artificial Analysis, Haiku 5.5 scores 43 points on the Intelligence Index, GPT-6 Luna 38 and Sonnet 5.5 56; on the same day Anthropic halved the cache read price for Sonnet 5.5 from $0.20 to $0.10 per million tokens, which is meant to make agentic tasks around 20 percent cheaper; Max and Team subscribers receive a monthly API credit of 00, 00 or up to $500 this week; for SMEs this means: classification, summaries, database queries and pre-sorting at volume become considerably cheaper; the extra consumption of the new tokenizer and the price step above 100,000 tokens belong in your own calculation, and the test figures come from the vendor.
EmbeddingGemma 2: Google releases a multimodal embedding model for on-device use under Apache 2.0
According to heise online on 7 October, Google has released EmbeddingGemma 2, an embedding model that maps text, code, images, audio and video into a shared vector space, so that video recordings can be searched with a spoken query, for example; the model builds on the Gemma 4 architecture and is modular: 270 million parameters for text and code, 170 million for the image encoder, 300 million for the audio encoder, 740 million when fully loaded; vectors have 768 dimensions by default and can be cut to 512, 256 or 128, and the context window is 8,192 tokens, four times that of the predecessor; according to Google, one input can hold about 5.5 minutes of audio, 29 images or 58 video frames; on a Pixel 11 Pro the quantised model needs about 191 MB of active RAM for the text weights and about 567 MB for the full multimodal variant, by Google's account; on the MTEB Code benchmark the score rises from 68.76 to 78.68 points, again by Google's figures; the weights are on Hugging Face and Kaggle under Apache 2.0, and Google names LiteRT, MediaPipe, Transformers and llama.cpp as integrations; Simon Willison noted on 6 October that open weights matter especially for embedding models, because stored vectors have to be recomputed when the model changes; the article does not name supported languages; for SMEs this means: anyone building a local search across documents, images or recordings gets a small model with a free licence that runs on the device and keeps data in house; the scores come from Google, so quality on your own data should be checked before deployment.
Google opens the SynthID detector to anyone with a Google, Apple or ChatGPT account
According to heise online on 7 October, Google has widened access to its SynthID detector: the portal, which for some months was open only to registered media representatives, is now available to everyone by Google's account, provided they have a Google, Apple or ChatGPT account; the tool checks whether an image, a video or an audio file carries a SynthID watermark, which indicates AI generation; it recognises content from Google's own models and from the partners OpenAI, Nvidia and Kakao, with Apple to follow soon; Google has embedded the watermark, invisible to humans, since 2023 when generating media, and it can also be read via Google Search, the Gemini app and Chrome; anyone who declines to register for privacy reasons is left out, heise has asked Google whether that is to change, and no answer had arrived at publication; third-party services also read SynthID, on other terms and at other costs; only content that carries a SynthID watermark is detected; the article names neither prices nor countries; for SMEs this means: anyone who wants to check incoming images, videos or audio files for AI origin, for instance in job applications, damage claims or supplier documents, now has a tool without special access; a missing watermark proves nothing, because only content from the participating providers is marked.
Meta and Sierra present the Personal Agent Protocol: companies decide how AI agents use their website
According to heise online on 7 October, Meta and the AI service provider Sierra have presented the Personal Agent Protocol, a standard for how AI agents interact with company websites on behalf of users, for example to make purchases; companies decide how an agent may use their site: directly via the website, via an API based on a standard such as Model Context Protocol or OpenAPI, or via the company's own AI agent; they can also observe what agents do on their site; customers who send out an agent choose whether it gets read access only, to gather information, or write access as well, to carry out actions, and sessions are built on OAuth; launch partners are Genesys, Instinct, Rocket, Shopify, Stripe and Walmart; Visa is pursuing a competing standard, and Stripe and Shopify are on board with both; the specification is not yet public and, according to the article, is to appear as version 0.1 in the course of the month; Meta's recently introduced personal agent Muse is to shop online and carry out actions for users in future; the article names no licence; for SMEs this means: anyone running an online shop or customer portal will have to decide in the coming months whether and how third-party agents may use the site; two competing standards are taking shape, so before committing it is worth waiting for version 0.1 and for the position of your own shop or payment provider.
Liquid AI releases open d1 decision models: one forward pass per answer, up to 50 milliseconds on edge devices
According to Liquid AI on 7 October on the Hugging Face blog, two decision models from the d1 family have been released with open weights; unlike generative models they produce no tokens; they answer a question with predefined answer options in a single forward pass; d1-3B with 3 billion parameters is based on LFM2.5-VL-3B and accepts text and images, d1-omni-600M with 0.6 billion parameters is based on LFM2.5-Encoder-350M and accepts text plus image or text plus audio, the latter as an early research release; across seven public datasets, including reading comprehension, toxicity, intent recognition and medical questions, d1-3B reaches a mean score of 82.9 against 81.1 for Decider 4B, and d1-omni-600M 78.4 against 77.1 for Decider 2B; on the Decision Index 0.2.1, d1-3B scores 48.57 points, ahead of Decider 35B-A3B at 47.11; d1-3B answers a single question in 8 milliseconds on an RTX 4090, in 30 on an Apple M5 Pro, in 16 on a Jetson AGX Thor and in 50 on a Jetson Orin Nano; both models are on Hugging Face, running them requires Transformers from version 5.14 and loading with the models' own code; the post names no licence, no image or audio benchmarks and no speed figures for the smaller model; for SMEs this means: for recurring yes-no or multiple-choice decisions, such as ticket routing, quality checks on the line or pre-sorting documents, there are now small models that run on an industrial PC without the cloud; before deployment, check the licence on the model card and the accuracy on your own data.
Hostinger tests the decision model Jev: about five times faster, 88.5 percent on phishing domains, too inaccurate for support chats
According to Hostinger on 7 October, Simon Lim and data scientist Mantas Lukauskas tested the model Jev from TypeSafe AI on two of their own tasks; Jev holds no conversation, it receives a question with fixed answer options and returns a probability per option plus a confidence score, and the vendor was co-founded by Diogo Almeida, formerly at OpenAI; on 781 previously unseen domains, Jev identified phishing in 88.5 percent of cases, the self-hosted Gemma 4 31B reached 90.0 percent and GPT-5.6 Luna, the best API model, 84.5; Jev took 0.38 seconds on average against 2.03 seconds for Gemma and cost about 0.021 dollars per 1,000 domains, about 4.3 times less than GPT-5.6 Luna; on 528 support chats, Jev assigned the right product in 79.3 percent of cases, GPT-5 mini in 86.9 percent, and on product plus intent together both landed at around 42 percent; median latency was 0.81 against 4.54 seconds, cost 0.23 against 1.27 dollars per 1,000 chats; Hostinger concludes that Jev suits high-volume checks where a small loss of accuracy is acceptable, while for product classification a gap of 7.6 points is too much to switch; the confidence threshold has to be calibrated per task on your own data and cannot be borrowed from other models; the model handles text only, does no multi-step planning and sees only what it is sent; for SMEs this means: for high-volume, clearly defined checks such as spam, phishing or duplicate detection, a specialised decision model can cut cost and waiting time considerably; the saving only counts if the hit rate holds, so measure on your own data first.
NTT Data opens an AI test lab in Munich: Mistral model, Red Hat OpenShift and hardware from Dell, Nvidia and AMD
According to heise online on 7 October, NTT Data has opened an AI lab in Munich-Lohhof where companies can test their AI applications under real hardware conditions before going into production, and Bavaria's digital minister Fabian Mehring attended the opening; the software on offer is the Mistral model with open weights and Red Hat's OpenShift platform; the infrastructure comes from Dell as an AI Factory, with Nvidia Blackwell systems in Dell Pro Max workstations, AMD Instinct MI355X and EPYC chips for open-source projects, and inference hardware from Axelera Europa in Dell Pro Slim Plus machines, and all parts can be swapped during a test phase; customers can also bring their own environments or install their own racks and use only power and cooling; by the provider's account, full sovereignty over data and applications stays with the customers, and access is either via the internet or fully isolated from the network; NTT Data also offers consulting, metrics, reference architectures and AI agents for provisioning and operations; the article names neither prices, nor an application process, nor capacities; for SMEs this means: anyone who wants to run an AI application themselves and does not yet know which hardware it needs can measure it here before investing; what the lab costs and whether it is open to small firms is a question for the provider.
Everlast AI: China's industrial AI relies on small models instead of language models, humanoids take on logistics first
According to the newsletter Everlast AI on 7 October, Benny Drescher, head of technology at the Hong Kong Industrial Artificial Intelligence and Robotics Centre (FLAIR), gives his assessment of China's industrial AI in an interview; FLAIR was set up as a joint project of the Hong Kong Productivity Council and the RWTH Aachen Campus, and Drescher's team trains its own AI models for plants in southern China; at the display maker TCL an AI station at the end of the line inspects every panel, detects the defect, assigns it to a class and delivers the cause, and the line adjusts its process parameters until the fault becomes rarer; large language models, Drescher says, are not enough for industrial AI, because sensor readings and CAD data count on the shop floor, and they appear in no public training set, and because quality inspection demands accuracy above 95 percent; he therefore trains small, specific models on plant data that run on the plant's own hardware on site; for humanoids he sees the first solid use case in logistics, AgiBot and UBTECH put their robots on wheels instead of legs, and final assembly still lacks the positioning accuracy; another model from his team computes a collision analysis and the optimal assembly sequence from CAD data; German mid-sized firms with plants in China tell him a prototype takes a month there and a year in Germany; the newsletter recommends small, measurable pilot projects on a single line to manufacturing clients; for SMEs this means: the first AI step in manufacturing is often a small, in-house model for quality inspection or root-cause analysis, trained on plant data and run on site; anyone evaluating humanoids starts with moving crates and not yet with assembly.
Grok Bot becomes a model router: SpaceX reportedly plans to use Claude, Midjourney and Suno too
According to matching headlines from Gizmodo, The Next Web and The Information on 7 October, Grok Bot from Elon Musk's SpaceX is becoming a model router that also calls on models from competitors depending on the task, and the models named are Claude from Anthropic, Midjourney and Suno; The Information's headline quotes Musk as saying SpaceX will sometimes use rival models to power Grok Bot, and Gizmodo's headline places the step within SpaceX's continued shift toward AI infrastructure; so far only the headlines of the three outlets are known, so selection rules, prices and data flows remain open; for SMEs this means: even a vendor with its own frontier models no longer ties itself to a single model; anyone buying AI services asks the vendor which models sit behind an answer and where the data ends up in the process.
Compiled automatically · always up to date · summaries may be translated — the linked source can be in another language.
Frequently asked questions about AI for mid-sized companies
What does iConference AG do?
We are an AI consultancy based in Zug, working with SMEs and mid-sized companies in Switzerland and Europe. We start with the foundation of your business: processes, data, systems and compliance. The technology builds on that, with RAG on your own knowledge. Two areas: consulting from assessment to production use, and AI services, where we handle individual tasks for you.
What are the four pillars of an AI deployment?
Processes and organisation, data hygiene, infrastructure and compliance. Tools only work as well as the foundation they stand on. A process that lives only in people's heads cannot be delegated, neither to people nor to AI. Scattered and outdated file storage is the quietest obstacle to AI. An honest assessment along these four pillars is the first step.
Where should a company start with AI?
With the foundation. The framework describes the technology, the foundation describes your business. Only once processes are documented, data is findable and the ground rules are settled will an AI tool carry weight. So the assessment comes first, then the choice of a use case with visible benefit.
What is an AI audit?
A structured assessment along four pillars: processes and organisation, data hygiene, infrastructure and compliance. We check which processes are documented and therefore delegable, where your data sits, whether your systems offer an export or an interface, and what ground rules apply to using AI. The result shows the largest opportunities and the open points.
What does an AI audit cost?
An AI audit combined with training starts at CHF 800, excl. VAT. This includes modules 1 and 2 of our workplace AI training, free of charge for two people. The audit delivers the assessment along the four pillars, the training brings your team to a shared level. A concrete quote follows after a short conversation.
Does our data stay in Switzerland?
Sensitive data stays in Switzerland. We use frontier models where their capability makes the difference, backed by anonymisation and clear guardrails. For data storage you choose between hosting in Switzerland, in the EU or worldwide. Which option fits depends on how protection-worthy your data is.
Which companies does iConference work with?
With SMEs and mid-sized companies in Switzerland and Europe. The approach is pragmatic and budget-aware: we work with what already exists in the business and aim AI use at measurable benefit. The company is based in Zug. We work on site in the canton of Zug, the Zurich area, Central Switzerland and the Aarau area, and by video elsewhere in Switzerland and in Europe.
What is the difference between AI consulting and AI services?
In consulting we enable your business to use AI productively on its own: audit, strategy, framework, training. In services we take on the task ourselves, for example revising contracts, restructuring product catalogues, turning documents into structured data or answering complex RFPs. For that you need no AI environment of your own.
Let's go. Bring AI into your business.
Tell us where you are today — we'll show you the next concrete step.
0,000 and is due to ship this year, but for most tasks it still needs a remote human operator; Unitree sells a humanoid for under $6,000; Elon Musk holds out the Optimus for the end of 2027 at about 0,000, after scaling back his goal of thousands of robots in Tesla factories by the end of 2025 to "some" doing simple tasks; an attempt by Google DeepMind to have a Gemini Robotics model pack the ingredients for a mushroom risotto into a basket failed; as reasons the article names a lack of training data from the physical world, the variety of real environments, unsolved navigation in new spaces and the requirement that a practical robot must almost always succeed, while many demonstrations are remote-controlled or scripted; Morgan Stanley nevertheless expects nearly one billion human-like robots by 2050 and a market above $5 trillion; for SMEs this means: for logistics with clearly defined grips, first trials are running in large companies, while for workshops, care or households no deployable solution is to be expected in the next few years; anyone reviewing offers should ask for the success rate without remote control.
Source: MIT Technology Review
Models7 Oct 2026
Claude Haiku 5.5: $0.10 per million input tokens up to 100,000 tokens, Sonnet 5.5 cache reads halved
According to Anthropic on 7 October, Claude Haiku 5.5 has been released, the new version of its smallest and cheapest model class; for prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, above that $0.50 and .50, while Haiku 4.5 was at and $5; Anthropic puts the saving at around 75 percent on average, heise online at up to 90 percent for short prompts; it is the first Haiku model with an adjustable effort setting from Low to Max, and the Python and TypeScript SDKs gain computer use and browser use in beta; it runs on the Claude Platform, AWS, Google Cloud and Microsoft Azure; on OSWorld 2.1 (offline subset) Anthropic reports 72.4 percent, on Humanity's Last Exam 45.9 percent without and 57.4 percent with tools; heise online reports a context window of one million tokens instead of 200,000 and up to 128,000 output tokens, which Anthropic's page does not state; the new tokenizer uses slightly more tokens per task according to Anthropic, heise puts it at around 30 percent more, Simon Willison measures around 25 percent more on a long prompt; according to heise, citing Artificial Analysis, Haiku 5.5 scores 43 points on the Intelligence Index, GPT-6 Luna 38 and Sonnet 5.5 56; on the same day Anthropic halved the cache read price for Sonnet 5.5 from $0.20 to $0.10 per million tokens, which is meant to make agentic tasks around 20 percent cheaper; Max and Team subscribers receive a monthly API credit of 00, 00 or up to $500 this week; for SMEs this means: classification, summaries, database queries and pre-sorting at volume become considerably cheaper; the extra consumption of the new tokenizer and the price step above 100,000 tokens belong in your own calculation, and the test figures come from the vendor.
EmbeddingGemma 2: Google releases a multimodal embedding model for on-device use under Apache 2.0
According to heise online on 7 October, Google has released EmbeddingGemma 2, an embedding model that maps text, code, images, audio and video into a shared vector space, so that video recordings can be searched with a spoken query, for example; the model builds on the Gemma 4 architecture and is modular: 270 million parameters for text and code, 170 million for the image encoder, 300 million for the audio encoder, 740 million when fully loaded; vectors have 768 dimensions by default and can be cut to 512, 256 or 128, and the context window is 8,192 tokens, four times that of the predecessor; according to Google, one input can hold about 5.5 minutes of audio, 29 images or 58 video frames; on a Pixel 11 Pro the quantised model needs about 191 MB of active RAM for the text weights and about 567 MB for the full multimodal variant, by Google's account; on the MTEB Code benchmark the score rises from 68.76 to 78.68 points, again by Google's figures; the weights are on Hugging Face and Kaggle under Apache 2.0, and Google names LiteRT, MediaPipe, Transformers and llama.cpp as integrations; Simon Willison noted on 6 October that open weights matter especially for embedding models, because stored vectors have to be recomputed when the model changes; the article does not name supported languages; for SMEs this means: anyone building a local search across documents, images or recordings gets a small model with a free licence that runs on the device and keeps data in house; the scores come from Google, so quality on your own data should be checked before deployment.
Google opens the SynthID detector to anyone with a Google, Apple or ChatGPT account
According to heise online on 7 October, Google has widened access to its SynthID detector: the portal, which for some months was open only to registered media representatives, is now available to everyone by Google's account, provided they have a Google, Apple or ChatGPT account; the tool checks whether an image, a video or an audio file carries a SynthID watermark, which indicates AI generation; it recognises content from Google's own models and from the partners OpenAI, Nvidia and Kakao, with Apple to follow soon; Google has embedded the watermark, invisible to humans, since 2023 when generating media, and it can also be read via Google Search, the Gemini app and Chrome; anyone who declines to register for privacy reasons is left out, heise has asked Google whether that is to change, and no answer had arrived at publication; third-party services also read SynthID, on other terms and at other costs; only content that carries a SynthID watermark is detected; the article names neither prices nor countries; for SMEs this means: anyone who wants to check incoming images, videos or audio files for AI origin, for instance in job applications, damage claims or supplier documents, now has a tool without special access; a missing watermark proves nothing, because only content from the participating providers is marked.
Meta and Sierra present the Personal Agent Protocol: companies decide how AI agents use their website
According to heise online on 7 October, Meta and the AI service provider Sierra have presented the Personal Agent Protocol, a standard for how AI agents interact with company websites on behalf of users, for example to make purchases; companies decide how an agent may use their site: directly via the website, via an API based on a standard such as Model Context Protocol or OpenAPI, or via the company's own AI agent; they can also observe what agents do on their site; customers who send out an agent choose whether it gets read access only, to gather information, or write access as well, to carry out actions, and sessions are built on OAuth; launch partners are Genesys, Instinct, Rocket, Shopify, Stripe and Walmart; Visa is pursuing a competing standard, and Stripe and Shopify are on board with both; the specification is not yet public and, according to the article, is to appear as version 0.1 in the course of the month; Meta's recently introduced personal agent Muse is to shop online and carry out actions for users in future; the article names no licence; for SMEs this means: anyone running an online shop or customer portal will have to decide in the coming months whether and how third-party agents may use the site; two competing standards are taking shape, so before committing it is worth waiting for version 0.1 and for the position of your own shop or payment provider.
Liquid AI releases open d1 decision models: one forward pass per answer, up to 50 milliseconds on edge devices
According to Liquid AI on 7 October on the Hugging Face blog, two decision models from the d1 family have been released with open weights; unlike generative models they produce no tokens; they answer a question with predefined answer options in a single forward pass; d1-3B with 3 billion parameters is based on LFM2.5-VL-3B and accepts text and images, d1-omni-600M with 0.6 billion parameters is based on LFM2.5-Encoder-350M and accepts text plus image or text plus audio, the latter as an early research release; across seven public datasets, including reading comprehension, toxicity, intent recognition and medical questions, d1-3B reaches a mean score of 82.9 against 81.1 for Decider 4B, and d1-omni-600M 78.4 against 77.1 for Decider 2B; on the Decision Index 0.2.1, d1-3B scores 48.57 points, ahead of Decider 35B-A3B at 47.11; d1-3B answers a single question in 8 milliseconds on an RTX 4090, in 30 on an Apple M5 Pro, in 16 on a Jetson AGX Thor and in 50 on a Jetson Orin Nano; both models are on Hugging Face, running them requires Transformers from version 5.14 and loading with the models' own code; the post names no licence, no image or audio benchmarks and no speed figures for the smaller model; for SMEs this means: for recurring yes-no or multiple-choice decisions, such as ticket routing, quality checks on the line or pre-sorting documents, there are now small models that run on an industrial PC without the cloud; before deployment, check the licence on the model card and the accuracy on your own data.
Hostinger tests the decision model Jev: about five times faster, 88.5 percent on phishing domains, too inaccurate for support chats
According to Hostinger on 7 October, Simon Lim and data scientist Mantas Lukauskas tested the model Jev from TypeSafe AI on two of their own tasks; Jev holds no conversation, it receives a question with fixed answer options and returns a probability per option plus a confidence score, and the vendor was co-founded by Diogo Almeida, formerly at OpenAI; on 781 previously unseen domains, Jev identified phishing in 88.5 percent of cases, the self-hosted Gemma 4 31B reached 90.0 percent and GPT-5.6 Luna, the best API model, 84.5; Jev took 0.38 seconds on average against 2.03 seconds for Gemma and cost about 0.021 dollars per 1,000 domains, about 4.3 times less than GPT-5.6 Luna; on 528 support chats, Jev assigned the right product in 79.3 percent of cases, GPT-5 mini in 86.9 percent, and on product plus intent together both landed at around 42 percent; median latency was 0.81 against 4.54 seconds, cost 0.23 against 1.27 dollars per 1,000 chats; Hostinger concludes that Jev suits high-volume checks where a small loss of accuracy is acceptable, while for product classification a gap of 7.6 points is too much to switch; the confidence threshold has to be calibrated per task on your own data and cannot be borrowed from other models; the model handles text only, does no multi-step planning and sees only what it is sent; for SMEs this means: for high-volume, clearly defined checks such as spam, phishing or duplicate detection, a specialised decision model can cut cost and waiting time considerably; the saving only counts if the hit rate holds, so measure on your own data first.
NTT Data opens an AI test lab in Munich: Mistral model, Red Hat OpenShift and hardware from Dell, Nvidia and AMD
According to heise online on 7 October, NTT Data has opened an AI lab in Munich-Lohhof where companies can test their AI applications under real hardware conditions before going into production, and Bavaria's digital minister Fabian Mehring attended the opening; the software on offer is the Mistral model with open weights and Red Hat's OpenShift platform; the infrastructure comes from Dell as an AI Factory, with Nvidia Blackwell systems in Dell Pro Max workstations, AMD Instinct MI355X and EPYC chips for open-source projects, and inference hardware from Axelera Europa in Dell Pro Slim Plus machines, and all parts can be swapped during a test phase; customers can also bring their own environments or install their own racks and use only power and cooling; by the provider's account, full sovereignty over data and applications stays with the customers, and access is either via the internet or fully isolated from the network; NTT Data also offers consulting, metrics, reference architectures and AI agents for provisioning and operations; the article names neither prices, nor an application process, nor capacities; for SMEs this means: anyone who wants to run an AI application themselves and does not yet know which hardware it needs can measure it here before investing; what the lab costs and whether it is open to small firms is a question for the provider.
Everlast AI: China's industrial AI relies on small models instead of language models, humanoids take on logistics first
According to the newsletter Everlast AI on 7 October, Benny Drescher, head of technology at the Hong Kong Industrial Artificial Intelligence and Robotics Centre (FLAIR), gives his assessment of China's industrial AI in an interview; FLAIR was set up as a joint project of the Hong Kong Productivity Council and the RWTH Aachen Campus, and Drescher's team trains its own AI models for plants in southern China; at the display maker TCL an AI station at the end of the line inspects every panel, detects the defect, assigns it to a class and delivers the cause, and the line adjusts its process parameters until the fault becomes rarer; large language models, Drescher says, are not enough for industrial AI, because sensor readings and CAD data count on the shop floor, and they appear in no public training set, and because quality inspection demands accuracy above 95 percent; he therefore trains small, specific models on plant data that run on the plant's own hardware on site; for humanoids he sees the first solid use case in logistics, AgiBot and UBTECH put their robots on wheels instead of legs, and final assembly still lacks the positioning accuracy; another model from his team computes a collision analysis and the optimal assembly sequence from CAD data; German mid-sized firms with plants in China tell him a prototype takes a month there and a year in Germany; the newsletter recommends small, measurable pilot projects on a single line to manufacturing clients; for SMEs this means: the first AI step in manufacturing is often a small, in-house model for quality inspection or root-cause analysis, trained on plant data and run on site; anyone evaluating humanoids starts with moving crates and not yet with assembly.
Grok Bot becomes a model router: SpaceX reportedly plans to use Claude, Midjourney and Suno too
According to matching headlines from Gizmodo, The Next Web and The Information on 7 October, Grok Bot from Elon Musk's SpaceX is becoming a model router that also calls on models from competitors depending on the task, and the models named are Claude from Anthropic, Midjourney and Suno; The Information's headline quotes Musk as saying SpaceX will sometimes use rival models to power Grok Bot, and Gizmodo's headline places the step within SpaceX's continued shift toward AI infrastructure; so far only the headlines of the three outlets are known, so selection rules, prices and data flows remain open; for SMEs this means: even a vendor with its own frontier models no longer ties itself to a single model; anyone buying AI services asks the vendor which models sit behind an answer and where the data ends up in the process.
Compiled automatically · always up to date · summaries may be translated — the linked source can be in another language.
Frequently asked questions about AI for mid-sized companies
What does iConference AG do?
We are an AI consultancy based in Zug, working with SMEs and mid-sized companies in Switzerland and Europe. We start with the foundation of your business: processes, data, systems and compliance. The technology builds on that, with RAG on your own knowledge. Two areas: consulting from assessment to production use, and AI services, where we handle individual tasks for you.
What are the four pillars of an AI deployment?
Processes and organisation, data hygiene, infrastructure and compliance. Tools only work as well as the foundation they stand on. A process that lives only in people's heads cannot be delegated, neither to people nor to AI. Scattered and outdated file storage is the quietest obstacle to AI. An honest assessment along these four pillars is the first step.
Where should a company start with AI?
With the foundation. The framework describes the technology, the foundation describes your business. Only once processes are documented, data is findable and the ground rules are settled will an AI tool carry weight. So the assessment comes first, then the choice of a use case with visible benefit.
What is an AI audit?
A structured assessment along four pillars: processes and organisation, data hygiene, infrastructure and compliance. We check which processes are documented and therefore delegable, where your data sits, whether your systems offer an export or an interface, and what ground rules apply to using AI. The result shows the largest opportunities and the open points.
What does an AI audit cost?
An AI audit combined with training starts at CHF 800, excl. VAT. This includes modules 1 and 2 of our workplace AI training, free of charge for two people. The audit delivers the assessment along the four pillars, the training brings your team to a shared level. A concrete quote follows after a short conversation.
Does our data stay in Switzerland?
Sensitive data stays in Switzerland. We use frontier models where their capability makes the difference, backed by anonymisation and clear guardrails. For data storage you choose between hosting in Switzerland, in the EU or worldwide. Which option fits depends on how protection-worthy your data is.
Which companies does iConference work with?
With SMEs and mid-sized companies in Switzerland and Europe. The approach is pragmatic and budget-aware: we work with what already exists in the business and aim AI use at measurable benefit. The company is based in Zug. We work on site in the canton of Zug, the Zurich area, Central Switzerland and the Aarau area, and by video elsewhere in Switzerland and in Europe.
What is the difference between AI consulting and AI services?
In consulting we enable your business to use AI productively on its own: audit, strategy, framework, training. In services we take on the task ourselves, for example revising contracts, restructuring product catalogues, turning documents into structured data or answering complex RFPs. For that you need no AI environment of your own.
Let's go. Bring AI into your business.
Tell us where you are today — we'll show you the next concrete step.
