Mistral releases Mistral Large 4 open weights with a 1M-token context and a modified MIT license
Mistral Large 4 is now an open-weight model in practice and not only on paper. On October 29, 2026, Mistral published the weights of the trillion-parameter Mixture-of-Experts model it had introduced as a public preview on October 6, together with a post-preview version that the company says received three additional weeks of reinforcement learning. The checkpoints are on Hugging Face under a new license, the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT license, and the same model remains available through the Mistral API at the list price of $0.70 per million input tokens and $2.10 per million output tokens.
The release closes the gap left open at launch. When Mistral unveiled Large 4 at the start of the month, the model was usable only as a preview endpoint in Mistral Studio, and the announcement promised that the weights would follow at the end of October. That promise has now been kept, and the downloadable model is not a frozen copy of the preview: Mistral says it kept scaling reinforcement learning after the preview went live, with the largest gains in coding and defensive cybersecurity, and that those gains "continue to accelerate, with no signs of saturation". The company nicknames the model "Le Chonk", a name it already used at launch.
The architecture has not changed. According to the Hugging Face model card, Large 4 has about 1 trillion total parameters, 49 billion of which are active per token (52 billion when embeddings and output layers are counted, which explains the "A52B" in the repository name), a 1.6B-parameter Pixtral-style vision encoder and a 1M-token context window. It accepts text and images and produces text. What has changed is what developers can do with it: they can now download it, run it on their own hardware, fine-tune it and, under conditions discussed below, build commercial products on it.
··········
WHAT MISTRAL RELEASED ON OCTOBER 29.
The weights are published in the mistralai organization on Hugging Face as part of a Mistral Large 4 collection. The main repository, mistralai/Mistral-Large-4-1T-A52B, holds an FP8 checkpoint, and a second repository, mistralai/Mistral-Large-4-1T-A52B-NVFP4, holds a 4-bit NVFP4 quantization aimed at NVIDIA Blackwell GPUs, which Mistral describes as having a smaller memory footprint and faster inference. Hugging Face's launch post, published the same day, notes that no official BF16 checkpoint is listed; BF16 and lower-bit GGUF files are available from ggml-org and Unsloth for use with llama.cpp.
The release also includes a draft head for speculative decoding, which Mistral calls DSpark. The model card says it delivers up to 1.9x higher throughput than the target model alone, with the actual gain depending on hardware and workload. The figure is a throughput claim for the serving stack, not a statement about lower latency for every request, and Mistral does not publish the test configuration behind it.
On the architecture side, Hugging Face's technical breakdown fills in details that Mistral's own announcement leaves out. The text decoder has 72 layers with a hidden dimension of 6,144 and 64 attention heads. Attention uses Multi-head Latent Attention combined with a DeepSeek Sparse Attention indexer that selects up to 2,048 key positions per query, and KV state is shared across layers, two choices that directly target the memory cost of a 1M-token context. The first layer is dense; the other 71 layers are MoE layers in which a softmax router picks 4 of 128 routed experts per token, plus one shared expert. Positional encoding follows a 3:1 pattern of RoPE and NoPE layers.
........
Item | Mistral Large 4 open-weight release |
Total parameters | About 1T |
Active parameters | 49B per token (52B with embeddings and output layers) |
Experts | 128 routed, 4 active per token, plus 1 shared |
Vision encoder | 1.6B Pixtral-style encoder |
Context window | 1M tokens |
Official checkpoints | FP8 and NVFP4 |
Reasoning effort | low, medium, high (default high) |
License | Mistral Ecosystem Open-Weight License |
API model ID | mistral-large-4 |
........
The model is a hybrid instruct and reasoning system with three reasoning modes. Mistral says Low replaces the previous "None" setting and is meant for simple tasks, Medium for everyday work, and High, the default, for the hardest problems. Hugging Face notes that reasoning cannot be switched off entirely, so a request sent with reasoning_effort: none still runs at High. The model card recommends a temperature of 1.0 and top_p of 0.95 for every reasoning level.
··········
THE LICENSE IS OPEN WEIGHT, NOT OPEN SOURCE.
The Mistral Ecosystem Open-Weight License is the part of the release that matters most for companies deciding whether to build on Large 4. Mistral describes it as a modified MIT license that allows use, modification, fine-tuning and commercial use, with one explicit carve-out: a business that operates a Model-as-a-Service offering, in other words one that sells hosted access to the model to third parties, needs a separate commercial license from Mistral.
For most enterprise users, that condition does not get in the way. A bank running Large 4 on its own GPUs for internal agents, a software vendor embedding a fine-tuned version in an on-premise product, or a public administration deploying it in a sovereign data center all fall within the permitted uses as Mistral describes them. The restriction targets inference providers and cloud resellers that would otherwise compete directly with Mistral's own API using Mistral's weights. That also explains why the launch names specific partners, such as NVIDIA-hosted endpoints, OpenRouter and Microsoft Foundry, as distribution channels.
The distinction between open weights and open source applies here in full. Mistral publishes the trained parameters and inference support, but not the training data or the full training pipeline, and the license is not an OSI-approved open-source license because of the Model-as-a-Service clause. A full technical report is described as "coming soon", so details such as data composition and the reinforcement learning setup are still undisclosed.
··········
WHAT THREE MORE WEEKS OF REINFORCEMENT LEARNING CHANGED.
Mistral's announcement focuses on the difference between the October 6 preview and the released model. The clearest like-for-like comparison is on Terminal-Bench 4.0, an agentic benchmark of tasks carried out in a command-line environment: the released model scores 41%, up 13 percentage points from the preview's 28%, while producing 30% fewer output tokens on the benchmark. Mistral says this result leads open-weight models. On Lakera's B3 Agent Security Benchmark, which measures resistance to attacks against AI agents, the score rises from 93.3% to 97.1%.
........
Benchmark (Mistral-reported) | Preview or earlier checkpoint | Released model |
Terminal-Bench 4.0 | 28% (preview) | 41% |
B3 Agent Security, attack resistance | 93.3% (preview) | 97.1% |
Finch (FinWorkBench) | 67.4% (preview) | 73.3% |
BrowseComp | 75% (SFT checkpoint) | 84.5% |
Refusal rate on malicious cyber requests | 95% (preview, chart) | 96.6% average |
........
The BrowseComp figure needs a caveat that Mistral itself states: the 75% baseline is an earlier supervised fine-tuning checkpoint trained through a separate reinforcement learning pipeline, not the public preview, so the 9.5-point gain describes that pipeline rather than the difference users of the preview endpoint will notice.
On third-party indexes, Mistral reports 48% on the Artificial Analysis Coding Agent Index, which it calls effectively tied with Zhipu's GLM-5.3 while using 47% fewer tokens per task on average. The model card on Hugging Face adds a score of 26 on the Artificial Analysis Intelligence Index v5, which it describes as the highest among Western open-weight models. On a private benchmark, Cognition's FrontierCode 2.0, Mistral reports 19.7% and claims the lead among the models it compared. In Surge's blind pairwise human preference evaluation for code, Mistral says Large 4 was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash and placed behind only Claude Opus 5.5 among the models evaluated.
The October 6 announcement reported a "combined Coding Agent Index" of 49.8% for the preview. Mistral does not say whether that figure and the 48% now quoted come from the same index version, so the two numbers should not be read as a regression.
........
Benchmark (Mistral-reported) | Mistral Large 4 | GLM-5.3 | Kimi K3 | DeepSeek V4.1 Flash |
BrowseComp | 84.5% | 82.5% | 91.2% | 83.4% |
Finch (FinWorkBench) | 73.3% | 65.1% | 77.3% | Not reported |
B3 attack resistance | 97.1% | 93% | 88% | Not reported |
........
These comparisons are self-reported and selected by Mistral, and they show a mixed picture rather than across-the-board leadership: Moonshot's Kimi K3 stays ahead on BrowseComp and Finch, while Large 4 leads on agent attack resistance. For Finch, Mistral lists DeepSeek-V4-Pro-0813 at 67.4% rather than V4.1 Flash, and for B3 it lists DeepSeek-V4-Pro at 59%, so the DeepSeek comparisons are not like-for-like across rows.
··········
A CYBER DEFENSE MODEL WITH GUARDRAILS STILL BEING ADJUSTED.
Mistral positions the release as "built for coders, trusted by defenders", and the cybersecurity section of the announcement is unusually candid. The company reports an average refusal rate of 96.6% on malicious cyber requests, the highest it measures among open models, and says its guardrails are designed to distinguish defensive from offensive work so that the model avoids unnecessary refusals on legitimate tasks such as vulnerability analysis. On the Artificial Analysis Cyber Index the model scores 51%; the chart in the announcement shows competitor values between 34% and 57%, so Mistral does not claim the top position there.
The company also writes that, after monitoring live traffic, it observed the model refusing black-box attacks against live targets but sees "a potential for uplift to conduct real attacks". Its response is that it is "modifying our guardrails accordingly, and will continue updating them in the months to come". For an open-weight model this point has a specific consequence: guardrails Mistral updates on its own API do not automatically reach copies of the weights already downloaded and self-hosted, which is why responsibility for misuse controls shifts in part to whoever deploys the model.
Alongside the general release, Mistral is previewing a specialized cybersecurity variant of Large 4 for penetration testing, vulnerability assessment and red-teaming. It has no public name and is limited to enterprise design partners, with Orange Cyberdefense cited as one. Access is subject to what Mistral calls rigorous vetting and contractual safeguards, and testing is restricted to systems the partner owns or is explicitly authorized to test, in controlled environments. Booz Allen, Cisco and Cloudflare provided supporting statements; Booz Allen said the model "was able to match vulnerability detection recall rates against other frontier cyber models", without published figures behind the claim.
··········
HOW TO ACCESS MISTRAL LARGE 4 AND WHAT IT COSTS.
Developers now have three routes. The first is Mistral's own platform: Mistral Studio, the Mistral API under the model ID mistral-large-4, and Vibe Code, Mistral's coding environment. The second is partner hosting: OpenRouter, NVIDIA-hosted endpoints and an NVIDIA NIM container, and Microsoft Foundry, which Mistral says is the first cloud provider to bring Large 4 to customers, with private deployment options on Azure. The third is self-hosting from Hugging Face with vLLM, SGLang, Transformers or llama.cpp, with fine-tuning through TRL and PEFT or through Mistral Forge for enterprise customization.
API prices on Mistral's official pricing page, verified on October 30, 2026, are the same list prices announced for the preview. The 50% launch discount that the Mistral changelog announced for the first two weeks after October 6 has run its course on that schedule, and the full list prices now apply. The page also shows a batch option, but it does not list a separate batch price for Large 4, and no batch discount is stated.
........
Model (Mistral API) | Input | Cached input | Output |
Mistral Large 4 | $0.70 | $0.07 | $2.10 |
Mistral Large 3 | $0.50 | $0.05 | $1.50 |
Mistral Medium 3.5 | $1.50 | $0.15 | $7.50 |
Mistral Small 4 | $0.15 | $0.015 | $0.60 |
........
Prices are in USD per 1M tokens, as listed on the official Mistral pricing page on October 30, 2026. For a concrete sense of scale, take an internal assistant handling 10,000 requests a day for 30 days, with 3,000 input tokens and 800 output tokens per request and half of the input served from cache. On Large 4 that workload costs about $850 a month: $315 for uncached input, $31.50 for cached input and $504 for output. The same workload would cost about $608 on Large 3 and about $2,543 on Medium 3.5, which makes Large 4 roughly 40% more expensive than its predecessor on the API but well under half the price of Mistral's own mid-tier model on these assumptions. Output tokens account for most of the bill, so the 30% to 47% token reductions Mistral reports on agentic benchmarks would matter more to real costs than the per-token price, if they hold on production workloads.
··········
WHAT SELF-HOSTING A TRILLION-PARAMETER MODEL REQUIRES.
Running Large 4 locally is possible, but it is a data center workload. The model card's serving examples assume eight GPUs per node with tensor parallelism, and Hugging Face's Inference Endpoints deploy it on 8x NVIDIA H200. Fine-tuning the FP8 checkpoint with TRL uses one node of 8x B300, while the Dell Enterprise Hub supports H200, B200, GB200 and GB300 systems. The NVFP4 checkpoint requires NVIDIA Blackwell GPUs in vLLM. The model card lists the following recommended vLLM serving flags:
vllm serve mistralai/Mistral-Large-4-1T-A52B \
--tensor-parallel-size 8 \
--tool-call-parser mistral \
--reasoning-parser mistral \
--enable-auto-tool-choice \
--gpu-memory-utilization 0.85The Hugging Face team also tested aggressive quantization. With a 2-bit Q2_K GGUF, the model runs on a single Apple Silicon Ultra machine with 512 GB of unified memory, where the authors measured about 90 tokens per second of prefill and 15 tokens per second of generation on an M3 Ultra. On four RTX 6000 Pro cards with NVFP4 at a 32K context, they report 10 tokens per second of generation, falling to about 3 tokens per second with an IQ2_XXS quant at an 800K context. Those figures come from the authors' own tests and show that the 1M-token window is technically reachable on workstation hardware but impractical for interactive use there. On Inference Endpoints, a single request generated about 65 tokens per second, rising to 120 to 138 tokens per second on chat prompts where DSpark drafts were accepted.
For teams that want data to stay inside their own infrastructure, which is the main reason to choose an open-weight model of this size, the realistic options are an eight-GPU H200 or Blackwell node, a private Azure deployment through Microsoft Foundry, or Mistral's own European deployment, which the company has said it operates end to end. Readers who want the architecture background from the preview, including what the 1.05T total parameter figure does and does not mean, can find it in our earlier coverage of the Mistral Large 4 launch.
··········
WHAT IS CONFIRMED AND WHAT REMAINS OPEN.
What is established is that Mistral has delivered on its October 6 commitment: the weights of Mistral Large 4 are downloadable in FP8 and NVFP4, the license permits commercial use and fine-tuning except for Model-as-a-Service businesses, the 1M-token context and the 49B-active MoE design are documented in the model card, and the API keeps its $0.70 input and $2.10 output list prices. The released model is also measurably different from the preview on the benchmarks where Mistral gives both numbers, most clearly the 13-point gain on Terminal-Bench 4.0 and the higher attack resistance on B3.
What remains uncertain is equally concrete. Every benchmark result in the announcement is reported by Mistral or by its launch partners, and independent evaluations on the downloadable checkpoints have not yet been published. The technical report that would document training data and the reinforcement learning process has not been released. The model card on Mistral's documentation site still labels the API model as a public preview, so the terms of general availability, versioning and deprecation for mistral-large-4 are not yet clear. And Mistral's own statement that its cyber guardrails are still being modified means the safety behavior of the API model may diverge over time from that of self-hosted copies of the weights. For developers, the release makes Large 4 the most capable model Mistral has made available for local deployment; how it compares in production with Kimi K3, GLM-5.3 and the DeepSeek V4 family will depend on evaluations that have not been run yet.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
