Mistral opens weights for 1-trillion-parameter Large 4, still behind China overall

Mistral released open weights today for Mistral Large 4, the 1-trillion-parameter model it calls Le Chonk, and says three more weeks of reinforcement learning pushed it to the front of open-weight coding. The independent numbers back the coding claim in places and complicate it everywhere else: Artificial Analysis still ranks the model behind China's leading open models overall.

What three weeks bought

The model first appeared in public preview earlier this month. Mistral says it kept training for three weeks using reinforcement learning driven by customer and partner feedback. Large 4 is natively multimodal, taking text and images in and producing text and code, with 52 billion parameters active per step.

The headline gain is on Terminal-Bench 4.0, where Mistral says the score rose from 28% in preview to 41%. GLM-5.3 sits at 40%. Thirteen points over its own preview is real progress. One point over the next open model is a lead in the way a rounding error is a trend.

The Artificial Analysis Coding Agent Index is similar. Mistral reports 48%, which it calls effectively tied with GLM-5.3, while using 47% fewer tokens per task. On Cognition's private FrontierCode 2.0 benchmark, Large 4 leads open models at 19.7%. Claude Opus 5.5 scores 46% on the same test. Claude Code with Opus 5.5 tops the coding index at 62 points.

Human judges were kinder. In Surge Code's blind pairwise evaluations, Large 4 was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Against Opus 5.5 it won only 22% of matchups, according to The Decoder.

The overall index tells a different story

On Artificial Analysis Intelligence Index v5, Large 4 in high mode scores 26 points, up one from the preview's 25. So three weeks of training moved the general score by a single point. DeepSeek V4.1 Flash and GLM-5.3 each score 32, and MiMo-V2.6-Pro gets 30. Google's Gemini 4 Carbon leads at 59, one point ahead of Claude Opus 5.5.

Cost and speed improved. An index task now costs $2.06 on average, down from $5.54 for the preview. That is cheaper than GLM-5.3 ($5.01) and Kimi K3.1 Pro ($3.68), but far above DeepSeek V4.1 Flash ($0.60) and MiMo-V2.6-Pro ($0.26). Output tokens across the index fell from 150 million to 110 million, still slightly above the 100 million median for comparable models. Generation speed rose from 111 to 161 tokens per second.

Cyber is where the sources split. Mistral says Large 4 leads open models in defensive security. Artificial Analysis Cyber Index data cited by The Decoder puts it third among open models at 51, behind DeepSeek V4.1 Flash at 57 and MiMo-V2.6-Pro at 56. On CyberGym-E2E, which tests finding and patching memory-safety bugs in C and C++, Mistral reports 79%, matching MiMo-V2.6-Pro and beating DeepSeek V4.1 Flash (76%) and GLM-5.3 (29%). Pick the benchmark, pick the winner.

Mistral also says the model refused 96.6% of malicious cyber requests on average across JailbreakBench, StrongREJECT and AgentHarm. In the same announcement it writes: "we see a potential for uplift to conduct real attacks. We are modifying our guardrails accordingly." Vetted partners such as Orange Cyberdefense get a specialized variant for red-teaming and penetration testing.

What you actually get

The license is the Mistral Ecosystem Open-Weight license, a modified MIT license. The model is free to use, modify, fine-tune and commercialize, with one condition: businesses that operate a Model-as-a-Service offering need a separate commercial license from Mistral. How closely that fits the phrase "open weight" is a fair question. The announcement does not address it.

Hardware is not casual. According to SGLang documentation, the standard version runs on eight Nvidia B300 GPUs, and a more heavily compressed variant on four GB300s. The Next Web reports a quantized version for four B200 or B300 GPUs.

The API is priced at $0.70 per million input tokens, $0.07 cached input and $2.10 output. Preview list prices were $1.36 and $4.18. The weights are on Hugging Face, and the model is also available through Mistral Studio, the Mistral API, OpenRouter, Microsoft Foundry, Vibe Code and NVIDIA NIM containers.

The Next Web reports training on 3,800 Nvidia Grace Blackwell GPUs in Europe, with the reinforcement learning running on Mistral's own compute in France at about 10 megawatts. CEO Arthur Mensch framed the budget this way: "Other labs have raised hundreds of billions to develop equivalent capabilities; in three years, Mistral raised roughly $6 billion."

What to watch

The clock is already running. Moonshot plans to release weights for Kimi K3.1 Pro by November 4, 2026, and The Decoder notes that K3.1 Pro tops Terminal-Bench 4.0 at 67%. Mistral's 41% may hold its open-weight crown for about a week.

Next is the promised family of smaller models. Pierre Stock, Mistral's VP of Science, said: "There's going to be a family of models, and one of them is going to be on a laptop." Release dates, sizes and architectures are not given.

Then there is the evidence. FrontierCode 2.0 is private. Mistral says it assesses end-to-end code quality and whether a repository maintainer would merge the agent's change, but the announcement gives no further detail on the evaluation standards. The efficiency story, a trillion parameters for a fraction of rivals' funding, comes from the company, with no training-cost figures for the Chinese models alongside it.

The Deep View ties the enterprise pitch to Anthropic's IPO filing this week. Mistral is betting that buyers care about control and price as much as the top line of an index. The eval table says that bet is plausible. It does not yet say it is won.