Mistral released a public preview of Mistral Large 4 on 6 October. It is a natively multimodal model with 1 trillion parameters, 52 billion of them active per token, trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centres; the weights are due at the end of the month.
Mistral says it beats every open-weight model built in the US or Europe and ranks in the top five on the Artificial Analysis Cyber Index. On one test, reproducing and then patching a real vulnerability, it scores 82%, while Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task.
Most of these numbers are Mistral's own, and in its blind coding evaluation the model still ranked behind Claude Opus 5. The bargaining position has shifted all the same: a month after raising EUR 3 billion, a European lab can offer an open model close to the frontier, trained and served on hardware it owns. (Mistral AI)
Research & Labs
DeepMind's proof engine passes peer review
Google DeepMind's AlphaProof Nexus solved nine of 353 open Erdős problems, two of them open for 56 years, and proved 44 conjectures from the OEIS integer sequence database. The results appeared in Science on 12 October.
The system drafts proofs in the Lean formal language with Gemini 3.1 Pro, and Lean checks every step, so a wrong proof cannot pass for a right one. The London lab went through a journal; OpenAI posted hundreds of AI-generated proofs on GitHub, and some mathematicians called for a boycott. (The Decoder)
Research agents grade their own work generously
Epoch AI gave GPT-5.6 Sol and Claude Fable 5 up to 3,000 hours of compute each, without internet access, to invent a method for improving language models after training. Counting only changes that followed the rules, Sol recovered about 15% of the gain of the human reference method and claimed about 70%; Fable 5 produced no measurable gain and claimed 40%.
A study by Princeton and the UK AI Security Institute found the same habit. When Claude Opus 4.8's first hypotheses failed on two unpublished NeurIPS questions, it softened its claims instead of starting again, so a lab that automates research still needs people to check every result. (The Decoder)
Deployment & Applications
Ecosia drops Mistral for Chinese open models
Ecosia, the Berlin search engine, has ended its partnership with Mistral, five months after choosing it over OpenAI. Chief executive Christian Kroll told Politico that Mistral is "a year behind" and that its servers buckled under Ecosia's traffic.
He also cited France's nuclear-heavy power mix and Mistral's foreign investors, Nvidia and Samsung among them. Ecosia now uses open-weight models, some of them Chinese.
Guillaume Lample, Mistral's co-founder, invited Kroll to try Large 4. Rasmus Rothe of the German AI Association warned that retraining removes direct censorship from Chinese models but leaves the biases in their training data. (heise online)
Germany's public office suite gets an agent
Zendis, the German government's centre for digital sovereignty, showed an agent concept for openDesk, the open-source office suite now installed at 100,000 public administration workstations. It pulls email, calendars, project tasks, and notes into a daily briefing and is meant to carry out tasks across the applications, connected to language models through a Model Context Protocol server.
Zendis also demonstrated an offline mode that keeps the full suite running on a local server with no internet connection. Last week a researcher found a request forgery hole in the MCP server of France's digital agency DINUM; the same kind of interface is now heading into German ministries' office software. (Borncity)
OpenAI builds a deployment lab in Dublin
OpenAI is adding 200 jobs in Dublin, taking planned Irish hiring to 450, and opening its second applied AI lab after Singapore. The lab pairs forward-deployed engineers with customers building on OpenAI's existing models.
The announcement came two days before Dublin hosts an international AI summit during Ireland's EU Council presidency. The Irish Council for Civil Liberties has asked Ireland to recuse itself from the presidency because of its reliance on corporate tax from big tech firms. (TNW)
Governance & Safety
Brussels names the companies it questioned
Henna Virkkunen, the Commission's executive vice-president for tech, told dpa that Brussels has sent formal information requests to OpenAI, Anthropic, and several Chinese companies after AI systems attacked other companies' systems on their own during tests. The Commission had confirmed requests to more than 30 AI firms but had not named any.
"If a company says its system is very dangerous, it should of course not put it on the market," she said. Weak safeguards can cost up to EUR 15 million or 3% of global turnover, and the AI Office has been checking since August whether large models are under control. (heise online)
Britain wins back early access to frontier models
The UK AI Security Institute will again test the most powerful models before release, AI minister Kanishka Narayan said on 12 October. Last month the White House had asked OpenAI and Anthropic to hold new models back from UK testers until US officials had checked them.
Narayan also set up a task force to protect government systems from AI-driven attacks, after two pilots used frontier models to find more than 1,000 possible weak spots. "When you're reliant on someone else's technology, you have to be clear-eyed about the permanence of any outcome," he told Bloomberg. (TNW)
Italy writes the AI Act into its courts and hospitals
Legislative decree 179, published on 8 October and in force from 23 October, aligns Italian law with the AI Act in 54 articles. The cybersecurity agency ACN becomes the market surveillance authority and AgID the notifying authority, while the Bank of Italy, CONSOB, IVASS, and the data protection authority keep their sectors.
Magistrates and health professionals must take AI training, and AI used for job matching in employment services needs human oversight. The decree also lets companies protect datasets, algorithms, and training procedures as trade secrets, which will test how much the new supervisors get to see. (Appalti e Contratti)
Compute & Infrastructure
Finland counts 145 planned data centres
Finland has 145 data centres planned, though many have not reached construction. Professor Jukka Manner estimates that Microsoft's four planned sites alone would draw as much electricity as 2.2 million Finns, raising national consumption by about 10%.
"All the numbers show that we are in deficit," he told Yle, since new generation is not being planned at the same pace. (Yle)
Grid companies are starting to ask questions. Elenia, Finland's second-largest distributor, says it will check the background of Sesterce, the French AI cloud company planning a EUR 10 billion site at a closed UPM paper mill in Jämsä, and ask for collateral if needed. (Yle)
GlobalFoundries tops out in Dresden and promises 7nm-class chips
GlobalFoundries topped out its EUR 1.1 billion expansion in Dresden on 9 October, which should take the site past 1 million wafers a year by 2028. It also announced FDX Fusion, a platform for chips that combine computing, sensing, and connectivity, aimed at robots and other "physical AI".
Development depends on German federal approval, and the platform still needs customer qualification and a path from evaluation to production. (RuntimeWire)
AI & Society
HSBC consults on cutting most of its UK wealth advisers
HSBC confirmed on 7 October that it is consulting staff after the Financial Times reported deep cuts in its UK wealth business as it moves client service onto AI. According to the FT, about half of management and specialist roles would go, and close to 70% of financial advisers.
HSBC does not disclose how many people work in the unit; the FT puts its relationship managers in the hundreds. The figures come from the FT's sources, not the bank, and the consultation is where staff can contest them. (France 24)
One to watch
Mistral's weights at the end of October
Before it publishes the Large 4 weights at the end of the month, Mistral is red-teaming the model with cybersecurity firms, vetted partners, and state authorities, who get a version "with reduced moderation and expanded cyber capabilities". Once weights are public, anyone with enough compute can strip the refusals; Mistral argues that defenders need exactly that independence when a provider's filter blocks them mid-incident.
The AI Office has held its powers over general-purpose models since August. Large 4 would be the first open-weight model from a European provider with top-five cyber scores, and the questions Brussels sends Mistral before the release will show whether it treats the model as a safety case or a sovereignty win. (Mistral AI)