Mistral Large 4 goes open: a trillion parameters that tie GLM-5.3 on coding

One trillion parameters, a one-million-token context window and weights you can download since October 29, 2026: Mistral Large 4, nicknamed Le Chonk, is the largest model Mistral has built, trained entirely in Europe and licensed for commercial use. On coding it ties GLM-5.3, the best Chinese open model on that test, while burning almost half the tokens. Across the board, though, it still sits six points behind.

01What Large 4 brings

It's a multimodal Mixture-of-Experts model: text and images in, text out. Of its ~1T parameters, 49 billion are active per token (52 billion counting embeddings and output layers, hence the A52B in the checkpoint name). The router picks 4 of 128 experts per token, and one shared expert always runs.

The architecture stacks two kinds of sparsity: over experts (which feed-forward network handles each token) and over attention, with a DSA indexer that picks up to 2,048 earlier positions per query on top of Multi-head Latent Attention. GLM-5.3 and DeepSeek went the same way, and it's what makes a million-token context practical without the KV cache exploding. Mistral trained it from scratch on 3,800 Grace Blackwell GPUs in its own European data centers, and that is the pitch: "sovereign" AI you can run on your own servers.

The three weeks since the public preview weren't cosmetic. Reinforcement learning on tasks the team kept making harder lifted Terminal-Bench 4.0 from 28% to 41%, cut the average cost per Artificial Analysis index task from $5.54 to $2.06 and pushed output speed from 111 to 161 tokens per second.

02How to try it

On Mistral's API the model id is mistral-large-4. To host it yourself, the Hugging Face model card recommends vLLM's official image on an eight-GPU node:

docker run --runtime nvidia --gpus all \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -p 8000:8000 \
  --ipc=host \
  vllm/vllm-openai:mistral-large-4 \
  --model mistralai/Mistral-Large-4-1T-A52B \
  --tensor-parallel-size 8 \
  --tool-call-parser mistral \
  --reasoning-parser mistral \
  --enable-auto-tool-choice \
  --gpu-memory-utilization 0.85

The two --*-parser mistral flags aren't optional: without them tool calls break and reasoning isn't split from the answer. The server speaks the OpenAI-compatible API, so any client works:

curl -s http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"What is 17*23?"}],"max_tokens":512}'

Reasoning comes back in choices[0].message.reasoning and the answer in content. Keep in mind that max_completion_tokens counts reasoning tokens too.

Hardware is the real barrier. Per SGLang's docs, the standard build needs eight Nvidia B300s and the compressed one four GB300s. Hugging Face ships an Inference Endpoints recipe on 8× H200 with FP8 weights and 131K context, and Mistral offers an NVFP4 checkpoint for Blackwell. At the other end, the 2-bit GGUF fits on an Apple Silicon Ultra machine with 512 GB of unified memory: Hugging Face measured about 15 tokens/s on an M3 Ultra. Fine for tinkering, not for serving a team.

03Against the Chinese open models

The comparison that matters is with what you can download today. All figures come from Artificial Analysis, as reported by The Decoder:

On the Coding Agent Index, Large 4 scores 48%, effectively tied with GLM-5.3 while using 47% fewer tokens per task. For a closed reference point, Claude Code with Claude Opus 5.5 tops that index at 62, and on FrontierCode 2.0 Opus 5.5 reaches 46% against Large 4's 19.7%.

04Where it wins and where it loses

It wins on cheap, fast code. Matching GLM-5.3 with half the tokens and less than half the cost per task is what counts when your workload is a coding agent making hundreds of calls a day. In Surge's blind comparisons, raters preferred its code over GLM-5.3, Kimi K3 and DeepSeek more often than not (against Opus 5.5 it won just 22% of matchups).

It wins on control. Downloadable weights, a million-token context and a license that allows commercial use without asking. For a security team, not depending on a vendor's usage policy, which can cut you off in the middle of an incident, is a real argument.

It loses overall. Six points behind GLM-5.3 and DeepSeek V4.1 Flash on the main index, and the extra three weeks of RL only added one point. In cyber defense, which Mistral sells as its strength, it ranks third among open models (51 versus DeepSeek V4.1 Flash's 57) and costs more per task: $1.21 against a fraction of that for DeepSeek.

Its coding lead has an expiry date. Kimi K3.1 Pro hits 67% on Terminal-Bench 4.0, and Moonshot plans to release its weights by November 4, 2026. Mistral left it out of its charts because it isn't open yet; in a week it likely will be.

One license catch: if your company wants to sell Large 4 as a commercial cloud service, it needs a separate agreement with Mistral. Internal use, modification and shipping it inside your product are unrestricted.

05Verdict

Yes if you need a large model you can host in Europe or on your own infrastructure, your main workload is coding agents, and cost per task matters more than the last point of quality. Also if you're already on Mistral's API: at $0.70 / $2.10 its output costs a fifth of Kimi K3.1 Pro's.

No if you want the best open model overall (today that's DeepSeek V4.1 Flash and GLM-5.3, soon Kimi K3.1 Pro), or if you do security work on a budget. And without an eight-GPU node of current-generation hardware, the realistic way to use it is the API, which turns the open-weights advantage into a promise.

06Sources

• Paris-based Mistral releases "Le Chonk," a strong coding model that still trails China overall, The Decoder, October 29, 2026.
• Welcome Le Chonk: Mistral Large 4 on Hugging Face, Hugging Face.
• mistralai/Mistral-Large-4-1T-A52B, model card on Hugging Face.