Mistral AI published the full weights of Mistral Large 4 overnight NZ time, three weeks after the model entered public preview — and on schedule against the end-of-October release the company promised. The 1-trillion-parameter model, internally nicknamed Le Chonk, can now be downloaded, fine-tuned and self-hosted by anyone: the weights are live on Hugging Face as the Mistral-Large-4-1T-A52B checkpoint, it is served as mistral-large-4 through Mistral’s API, and Ollama’s library already lists it.
What actually shipped
The numbers behind the name: Large 4 is a natively multimodal mixture-of-experts model with 49 billion active parameters per token (52 billion counting embeddings and output layers), a 512K-token context window, and fluency in 160-plus languages including every official language of the European Union. Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacentres — roughly two months of training at around 10 megawatts, per the preview reporting we covered earlier — and the company leans hard on that lineage: “trained from scratch on European infrastructure,” full chain of custody from first token to release.
Three reasoning modes now ship with the weights — low, medium, and high, with high as the default — which amounts to a cost dial: developers can trade reasoning depth against token spend per workload rather than switching models.
The benchmark claims, and what sits underneath them
Per Mistral’s announcement, three weeks of additional reinforcement learning since the 6 October preview pushed Large 4 to the top of the open-weight coding pack: 19.7% on Cognition’s private FrontierCode 2.0 benchmark (measuring end-to-end code quality a repository maintainer would actually merge), 41% on Terminal-Bench 4.0 — up 13 points from preview — and 48% on the Artificial Analysis Coding Agent Index, effectively tied with the leading open-weight model GLM-5.3 while using 47% fewer tokens per task. In Surge Code’s blind pairwise evaluations, it was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5.
The security story is the sharper bet. Mistral reports 51% on the Artificial Analysis Cyber Index — which tests whether a model can find, reproduce and patch real vulnerabilities without breaking working software — and says Large 4 does this “without unnecessary refusals.” At preview, independent numbers collected by I-SCOOP had it at 50, level with GLM-5.3-Flash and behind Xiaomi’s MiMo-V2.6-Pro at 56, with Artificial Analysis expecting a top-three open-weight finish once weights went public. On the malicious side, Mistral claims an average 96.6% refusal rate on harmful cyber requests across JailbreakBench, StrongREJECT and AgentHarm — the highest it measures among open models.
Pricing on the API side, per I-SCOOP’s preview-era rundown: $1.36 per million input tokens and $4.18 per million output, with cached input at $0.14 — the discounted $0.68/$2.09 entry pricing we reported at preview tracking that list rate.
The security pitch has a built-in caveat — from Mistral itself
The commercial wedge is deliberate: Large 4 is being sold as the model that will do defensive security work that safety-filtered closed models refuse. Mistral is providing Orange Cyberdefense preview access to a specialised variant for penetration testing, vulnerability assessment and red-teaming, gated behind vetting and contractual safeguards and restricted to owned or explicitly authorised systems.
What makes the announcement unusual is that Mistral includes the counter-evidence in its own blog: monitoring live traffic since preview, its team observed the model refusing black-box attacks against live targets — but also saw “a potential for uplift to conduct real attacks,” with guardrail modifications underway. A model optimised to say yes to offensive-looking-but-defensive work sits closer to the misuse line than a refusal-heavy one; Mistral saying so out loud, in the same post that markets the capability, is worth more than the benchmark table. The independent re-run of the Cyber Index on public weights will settle whether 51% holds.
Why this one matters outside France
The open-weight race is no longer a China-versus-everybody story: Large 4’s competition on these boards is now GLM-5.3, Kimi K3, DeepSeek V4.1 Flash and MiMo-V2.6-Pro, and Mistral’s lead on several is a single point with a token-efficiency buffer — which is the part deployments actually feel, since inference cost scales with tokens, not with index rank. Our preview coverage framed the 27 October weights drop as the real test of Europe’s open-model thesis; the drop slipped a couple of days but landed, and it follows the same playbook as Mistral Medium 3.5 shipping with open weights — preview first, weights by month’s end, every time.
For a small market like New Zealand, the self-hosting option is the substance: a 512K-context frontier model that organisations can download, fine-tune on their own data and run inside their own perimeter is the difference between buying access to AI and owning capability — the sovereignty-by-download path that doesn’t require a national supercomputer, which is effectively the only path a country of five million has. Mistral’s backers just confirmed €3 billion of Series D funding to keep exactly this pipeline running; whether the free weights hold up against the closed frontier is now testable by anyone with $22,299 and a power outlet.
The honest read: this is a strong, verifiable increment — Europe’s biggest model, now inspectable — arriving in a field where “best open model outside China” may flip again within the benchmark cycle. That is how the open-weights race should work.
📰 Sources
- Mistral AI — Le Chonk: Built for Coders, Trusted by Defenders (29 October 2026)
- Hugging Face — mistralai/Mistral-Large-4-1T-A52B model card
- I-SCOOP — Mistral Large 4 brings trillion parameter open weights to cybersecurity and enterprise AI
- DQ India — Mistral AI challenges US AI giants with trillion-parameter model
- Ollama — mistral-large-4 library listing
