Mistral published the weights for Mistral Large 4 on October 29, roughly three weeks after the model first appeared in public preview. It carries a trillion parameters, was trained entirely in Europe, and ships with coding scores the company says no other open model matches.
Mistral Large 4 Drops the Preview Label
The French lab's internal nickname for the model is Le Chonk, internet slang for an extremely fat cat. Nothing about the deployment is fat. The preview opened on October 6, and the October 29 release is the part developers actually wait for: downloadable weights.
That distinction matters more than it sounds. A hosted preview lets you send prompts and read answers. Open weights let you run the thing on your own hardware, fine-tune it, and keep your data inside your own network. Mistral says the technical report is coming, which means the details researchers usually want, such as the full data mix, are still missing.
What the company did disclose is the shape of the model and the hardware behind it. Both numbers are large.
A Trillion Parameters That Only Wake 50 Billion
Mistral Large 4 uses a sparse mixture-of-experts architecture. Instead of pushing every token through the entire network, a router sends each token to a small subset of specialist sub-networks. Total size and working size become two different numbers.
Here they're very different. The model holds 1 trillion parameters in total but activates roughly 50 billion during inference. VentureBeat puts the active count at 49 billion; the preview announcement said 52 billion. Either way, you pay to run about 5% of the model on any given token.
Training ran from scratch in Mistral's own European data centers on about 4,000 Nvidia Grace Blackwell GPUs over roughly two months, according to VentureBeat. The previous generation, Large 3, was 675 billion parameters with 41 billion active, trained on 3,000 H200s. Language coverage runs past 160 languages, including every official language of the European Union.
Input is natively multimodal, so text and images both go in. Output is text only. If you were hoping for image generation, that isn't what this is.
Mistral also split reasoning into three gears: Low for simple tasks, Medium for everyday work, and High as the default for the hardest problems. Low replaces the old no-reasoning setting rather than sitting alongside it.
The Coding Scores, All Company Reported
Every number in this section comes from Mistral. Some sit on third-party leaderboards, but Mistral chose which ones to cite and how to frame them.
On FrontierCode 2.0, a private benchmark from Cognition that grades end-to-end code quality and whether a repository maintainer would actually merge the result, Mistral Large 4 scores 19.7%. That's the highest figure Mistral lists for an open-weight model. DeepSeek V4.1 Flash follows at 19.3%, then GLM-5.3 at 17.8%, Kimi K3 at 17.3%, GLM-5.3 Flash at 14.2%, and Beam 1.0 at 2.9%.
Terminal-Bench 4.0 lands at 41%, a 13-point gain over the preview build, with 30% fewer output tokens. On the Artificial Analysis Coding Agent Index it reaches 48%, roughly level with GLM-5.3, while using 47% fewer tokens per task on average.
VentureBeat added a set of early figures from Mistral: DeepSWE v1.1 at 62%, against Beam at 44%, Qwen 3.8 Max at 51%, DeepSeek V4 Pro 0813 at 57%, and GLM-5.3 at 61%. The same report cautions that each vendor runs these evaluations under different conditions, so cross-company comparisons come with slack built in.
Where It Still Loses, by a Lot
The most interesting result here isn't a win. On Surge Code, a blind human preference test, Mistral Large 4 beats GLM-5.3 in 64% of comparisons, Kimi K3 in 56%, and DeepSeek V4.1 Flash in 78%. Against Claude Opus 5.5 it wins 22%, which means it loses roughly four times out of five.
That gap deserves a second look. Open weights have closed much of the distance on scripted benchmarks, where a fixed test harness rewards a particular kind of competence. They haven't closed the distance on the messier question of which output a person would rather keep.
None of this makes the model weak. A 22% win rate against a frontier closed model is a real result for something you can download and host tonight. It does mean the open models caught up framing needs a qualifier attached.
Security Work, as Described by the Vendor
Mistral's pitch leans hard on defensive security, which is an unusual angle for a model release.
On the Artificial Analysis Cyber Index, which tests finding vulnerabilities in a codebase, reproducing them, and patching without breaking working code, the company reports 51%. It reports refusing 96.6% of malicious cyber requests drawn from cyber-related prompts in JailbreakBench, StrongREJECT, and AgentHarm, and calls that the highest refusal rate it has measured on an open model. On the B3 agent security benchmark it resists 97.1% of attacks, up from 93.3% in preview.
Two more to round it out. Finch, a finance benchmark, comes in at 73.3%, second among open models behind Kimi K3 at 77.3%. BrowseComp, which measures hard web research, reaches 84.5%.
Mistral's own risk note is more candid than most. It says monitoring of live traffic shows the model refusing black-box attacks on real targets, while acknowledging potential uplift for people attempting actual attacks. The company says it's adjusting guardrails. It also offers a separately tuned cyber variant to design partners such as Orange Cyberdefense for penetration testing, vulnerability assessment, and red teaming, under strict vetting, contractual limits, and a rule that customers only test systems they own or are authorized to touch.
Open Weights, One Condition
The license is Mistral's own, not Apache or MIT. It comes with a single additional condition: if you offer the model to others as a managed service, meaning model as a service, you need a separate commercial license from Mistral.
For a research team, a startup running it internally, or a company deploying on its own infrastructure, that condition doesn't bite. For a cloud provider planning to resell inference, it does. Read the license before you build a business on top of it.
Weights are on Hugging Face, with vLLM or SGLang for self-hosting. Managed access runs through Mistral Studio, the Mistral API, OpenRouter, NVIDIA hosted endpoints, and Vibe Code for testing.
API pricing is $0.70 per million input tokens, $0.07 for cached input, and $2.10 per million output.
What the Release Actually Changes
A trillion-parameter model with open weights would have been unthinkable two years ago, and the reason it's possible now has less to do with ambition than with routing: activating 50 billion parameters is a bill most infrastructure teams can reason about.
The plain read is that Mistral built a strong coding model, priced it aggressively, and handed out the weights, while the strongest closed systems still win on human preference. That's a good trade for anyone who needs to keep code inside their own walls, and a warning for anyone who assumed open weights had already won.
