Mistral released the weights of Mistral Large 4 on 29 October, a one-trillion-parameter multimodal model it nicknames Le Chonk. It is the open release the company flagged in early October, after three weeks of extra reinforcement learning on the public preview.
Mistral says the model now leads open-weight models on Cognition's FrontierCode 2.0 at 19.7%, just ahead of DeepSeek V4.1 Flash at 19.3%. It ties GLM-5.3 on the Artificial Analysis Coding Agent Index at 48% while using 47% fewer tokens per task. These are Mistral's own figures.
The sovereignty pitch is the training record: Mistral says it trained the model from scratch on European infrastructure, so a buyer can audit its lineage and host it themselves. That is Mistral's answer to the question Berlin was debating the same week. (Mistral)
Deployment & Applications
DeepMind's planning tool halves officer time in three councils
Google DeepMind published pilot results on 30 October for PlanAssist, a Gemini-based tool that checks householder planning applications and drafts cited reports. Across Barnet, Camden and Dorset, the average time an officer spent on those tasks fell 58%.
More than 70% of the generated reports were approved with no edits. Officers keep the final say on every application.
Householder applications can make up 70% of a planning officer's workload, and the UK wants 1.5 million new homes by 2029. The pilot began in June, and the figures are DeepMind's own. (Google DeepMind)
Governance & Safety
Berlin asks whether to build AI or apply it
At the Digitalgipfel opening on 2 November, digital minister Karsten Wildberger said Germany is "still too slow" and wants industrial AI for the Mittelstand. Siemens chief Roland Busch said the race for large general-purpose models is largely lost to the US and China, and the open question is how to build AI into design and production.
Roberto Viola, the Commission's digital director-general, said the Digital Omnibus is folding ten data laws into one rulebook, including easier GDPR-compatible AI training. He said Brussels will revisit overlapping regulatory layers in the coming months, and that simplification does not replace investment in AI capacity.
Civil society groups repeated that the ministry works as a digital economy ministry. The programme leaves little room for open discussion with critics. (heise online)
Google will honour one shared AI-training opt-out from 2027
From 2027 Google plans to read a single policy token in robots.txt that tells all participating AI developers a site does not allow training use. The token comes from the IETF's AI Preferences working group.
Today a site operator lists each vendor's crawler separately, such as GPTBot, ClaudeBot or Google-Extended. Google-Extended stays for site-specific exceptions.
Compliance remains voluntary, and some crawlers already ignore robots.txt. A publisher gets one line to maintain instead of a growing list, but only vendors that adopt the token will read it. (heise online)
Compute & Infrastructure
Canada joins EuroHPC as its 39th member
Canada became the 39th participating state of the EuroHPC Joint Undertaking on 30 October, at the EU-Canada Summit in Montréal. The two sides have cooperated on research since 1996, and Canada has been associated with Horizon Europe since 2024.
The press release frames the move as a step towards a deeper relationship that pools partners behind Europe's supercomputing capacity. A new member adds partners, not capacity, so the gigafactory selection due in early 2027 will say more about Europe's compute than the new flag. (EuroHPC JU)
AI & Society
SUSE cuts a tenth of its staff and cites AI
SUSE, the Nuremberg Linux company owned by EQT, is cutting about 10% of its workforce. In a memo seen by The Register, chief executive Dirk-Peter van Leeuwen wrote that "automation and AI are fundamentally shifting how software companies operate" and that the requirement for some roles "has greatly diminished".
With more than 1,900 staff, that is roughly 190 people. Affected employees hear by 4 November, employee representatives are nominated the next day, and the strategy for fiscal 2027 follows on 16 November.
TNW notes other employers tying cuts to AI in recent weeks, including the Netherlands with 25,000 government jobs. This case has no named roles yet, so the consultation with staff representatives is where the detail will appear. (TNW)
One to watch
Poland's AI commission and its first complaints
Complaints filed with Poland's new AI commission can now trigger inspections of companies, Rzeczpospolita reports. Experts it spoke to say many firms are not ready.
The test is whether the commission opens its first inspections over the next six months, and whether it publishes guidance before it fines anyone. (Rzeczpospolita)