Mistral Large 4 'Le Chonk': Trillion Parameters, Open Weights
Mistral AI has published the weights of Mistral Large 4, its first trillion-parameter model. Nicknamed Le Chonk internally, it is a granular mixture-of-experts system with 1.05 trillion total parameters, about 49 to 52 billion active per token, a 1-million-token context window and image inputs. One checkpoint handles both quick chat questions and long agentic runs, switched by a reasoning effort setting with low, medium and high levels. The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own data centers in Europe, and it is already available in the vega.chat model picker and through api.vega.chat.
Two releases, one month apart
Mistral's model registry lists two Large 4 versions: a public preview from October 6, 2026, now marked retired, and the final checkpoint from October 29, 2026, which is the active release. The open weights on Hugging Face correspond to the October 29 version, improved with extra RL training. Figures quoted during the preview period, including the original API pricing, describe the retired checkpoint rather than the shipping one. Even the official documentation page still carries the October 6 date and a Public Preview label.
The previous flagship, Mistral Large 3, arrived in December 2025 as part of the Mistral 3 family. Ten months between flagships is a fast cadence at this size.
Architecture: from 675B to one trillion
| Mistral Large 3 | Mistral Large 4 | |
|---|---|---|
| Total parameters | 675B | 1.05T |
| Active per token | 41B | ~49B (52B with embeddings and output layers) |
| Architecture | Granular MoE | Granular MoE |
| Vision encoder | 2.5B | 1.6B Pixtral-style |
| Context window | 256k | 1M tokens |
| Reasoning | none | low / medium / high, high by default |
| Training cluster | 3,000 NVIDIA H200 | 3,800 NVIDIA Grace Blackwell |
| License | Apache 2.0 | Mistral Ecosystem Open-Weight License |
According to the vLLM team's day-0 writeup, the trunk has 72 MoE layers with 128 routed experts (top-4) plus one shared expert and a hidden size of 6,144. Attention is MLA with a sparse DSA indexer that selects the top 2,048 keys per query, part of how the million-token context stays tractable. The Pixtral-style encoder accepts images of different sizes and aspect ratios, several per request. Official checkpoints ship in FP8 and NVFP4, both using an FP8 latent KV cache. A built-in DSpark drafter runs speculative decoding on three sliding-window MLA layers, accepts about 2.9 tokens per step, and delivers close to a 2x throughput gain on TP8.
Sparse activation does not make this a small model in disguise. Only ~49B parameters fire per token, but storing 1.05T weights plus the KV cache still takes a serious multi-GPU node.
Benchmarks: strong claims, mixed receipts
Mistral's reported numbers for the final checkpoint look strong on paper:
| Benchmark | Result | Note |
|---|---|---|
| FrontierCode 2.0 | 19.7%, first place | Cognition's closed end-to-end coding eval |
| AA Coding Agent Index | 48% | Roughly GLM-5.3 level on 47% fewer tokens per task |
| Terminal-Bench 4.0 | 41% | +13 points over the preview, open-weight lead |
| Cyber Index | 51% | Finds, reproduces and patches vulnerabilities |
| Malicious cyber request refusals | 96.6% | Highest among open models, per Mistral |
| B3 Agent Security | 97.1% | Up from 93.3% in the preview |
| Finch (finance, tables) | 73.3% | Second among open weights after Kimi K3 |
| BrowseComp | 84.5% | Above GLM-5.3 and DeepSeek Flash 4.1 |
| Surge Code (blind human prefs) | Beats GLM-5.3, Kimi K3, DeepSeek V4.1 Flash | Behind only Claude Opus 5.5 |
The independent lab Vals AI measured the same open checkpoint and got different numbers: 26.26% on Terminal-Bench 4.0 against Mistral's 41%, 81.56% on Vibe Code Bench, 61.72% on a finance agent eval, and a 56.42% Vals Index that places the model 17th out of 30. A 15-point gap on an identical benchmark name usually means the harness, tools, token budget and stop criteria differed between runs. When you compare models, compare configurations.
A hands-on test by MindStudio adds a practical caveat. Across eight agentic coding tasks the model scored 42 of 80: a local Gemma 2B fine-tune earned 10 out of 10 and a three-elevator simulation 9.5, but two tasks scored zero because the model ended its turn without creating the file or returning an answer at all. The weak point is reliability: in those runs the model stopped before producing anything, and it can still return a confident summary while the artifact underneath is broken or missing.
Pricing
Final-release API pricing is $0.70 per million input tokens and $2.10 per million output, with cached input at $0.07. That is 40% above Large 3 per uncached token, though well below the preview's $1.36 input and $4.18 output.
| Large 3 | Large 4 | |
|---|---|---|
| Input per 1M tokens | $0.50 | $0.70 |
| Cached input | ~90% discount | $0.07 |
| Output per 1M tokens | $1.50 | $2.10 |
Mistral's counterargument is efficiency: if the model finishes agentic tasks on 30 to 47% fewer tokens than rivals, cost per completed task matters more than cost per million tokens. Worth measuring on your own workload before drawing conclusions either way.
The license asterisk
Large 4 ships under the Mistral Ecosystem Open-Weight License, a modified MIT. Use, modification, fine-tuning, distribution and commercialization are all allowed, with one condition: a company running a Model-as-a-Service business, where third parties get inference or post-training access with meaningful control over inputs and parameters, needs a separate commercial license from Mistral. Internal use and use through Mistral's own products are exempt. Large 3's Apache 2.0 has no such clause, which makes it the legally simpler option for hosting providers.
Deployment notes
vLLM shipped day-0 support with a substantial engineering list. NVFP4 runs on four Blackwell GPUs (or TP8 on eight). Decode context parallelism shards the latent KV cache across ranks for about 7.9x more KV capacity on TP8, and HiSparse offloads KV to host memory, growing a GB300 node's KV capacity from 1.97M to between 8.77M and 11.25M tokens. SGLang, Transformers and Docker Model Runner are supported too, and Microsoft Foundry offers private Azure deployments for organizations that want the model on their own infrastructure.
One integration trap from the ecosystem: the Vercel AI SDK's Mistral provider silently dropped the reasoning_effort parameter for Large 4 because the model id was missing from a hardcoded allowlist, so requests went out with no reasoning at all. The bug is fixed, with regression tests. The lesson generalizes: verify that reasoning and tool parameters actually reach the model, because "the SDK supports Mistral" says nothing about that.
Safety, with open caveats
Mistral gave safety unusual prominence in this release: the 96.6% refusal rate on malicious cyber requests, 97.1% attack resistance on the B3 Agent Security Benchmark, and a specialized cyber variant for pentesting and red-teaming handed to partners like Orange Cyberdefense under strict contracts. The company also states plainly that live-traffic monitoring showed potential for uplift in real attacks, so the guardrails keep evolving. The technical documentation warns that the model is not built to be infallible and is not meant for critical decisions without validation and human oversight. Strong safety metrics and unfinished agent tasks coexist in the same model, as the MindStudio results show.
Large 3 or Large 4
Pick Large 4 when you need the 1M context, reasoning modes, or the strongest open-weight numbers in agentic coding and cyber tasks, and when your deployment can handle a trillion-parameter checkpoint. Stay on Large 3 when price per token or license simplicity matter more: Apache 2.0 imposes no MaaS clause, the uncached token costs 40% less, and the card documents single-node deployment on A100 or H100 in NVFP4. Large 3 has its own known weaknesses, including NVFP4 quality drops past 64k context and shakier tool calling in independent tests.
Whichever you pick, run a pilot on the October 29 checkpoint with your real tasks, check that tool calls fire and artifacts actually get written, and pin the version and sampling configuration next to every number you record. Mistral's recommended settings are temperature 1.0 and top_p 0.95.
Mistral Large 4 is live in the vega.chat model list, and developers can call it through the api.vega.chat compatible API.