# Mistral Large 4

> Source: https://aiwiki.ai/wiki/mistral_large_4
> Updated: 2026-11-01
> Fact-checked: 2026-10-30
> Categories: AI Models, Large Language Models, Mixture of Experts, Multimodal AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Mistral Large 4." aiwiki.ai, 1 Nov 2026. https://aiwiki.ai/wiki/mistral_large_4
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Mistral Large 4 is a general-purpose [multimodal model](https://aiwiki.ai/wiki/vision_language_model) developed by [Mistral AI](https://aiwiki.ai/wiki/mistral_ai). It entered public preview on October 6, 2026, with a [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) architecture and API access.[1] On October 29, 2026, Mistral released the weights of an improved version under the Mistral Ecosystem Open-Weight License, a modified MIT license, and cut the API price to $0.70 per million input tokens and $2.10 per million output tokens.[18] The released checkpoint has 72 layers and about 1 trillion total parameters, of which 49 billion are active per token (52 billion counting embeddings and output layers), and is published in [FP8](https://aiwiki.ai/wiki/fp8) and [NVFP4](https://aiwiki.ai/wiki/nvfp4) versions.[17][23][24] It belongs to the [Mistral Large](https://aiwiki.ai/wiki/mistral_large) family, but its specifications and license should not be confused with those of [Mistral Large 3](https://aiwiki.ai/wiki/mistral_large_3).

<figure class="w-40">
<img src="/images/wd-2026-10-28-mistral_ai-logo-c17a15515f7e-tile.png" alt="Mistral AI logo" width="480" height="359" />
<figcaption>Mistral AI logo. Source: <a href="https://commons.wikimedia.org/wiki/File:Mistral_AI_logo_(2025%E2%80%93).svg">Wikimedia Commons</a>, <a href="https://commons.wikimedia.org/wiki/File:Mistral_AI_logo_(2025%E2%80%93).svg#Licensing">Public domain</a>.</figcaption>
</figure>

## Release and availability

Mistral called the model ML4 and "le Chonk", offered a Studio preview API, and announced weights for the end of October 2026.[2]

### Open-weight release

On October 29, 2026, Mistral released the weights on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) as Mistral-Large-4-1T-A52B.[17][18] The released model is not the preview checkpoint: Mistral says three more weeks of [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) improved its coding and defensive cybersecurity results and its reasoning efficiency.[18] It now offers three reasoning modes, low, medium and high, with high as the default; low replaces the earlier "None" mode.[18]

The weights use the Mistral Ecosystem Open-Weight License, a modified MIT license that permits commercial and non-commercial use. A company or its affiliates that operates a "Model as a Service" business must sign a separate commercial license with Mistral before using the model or its derivatives commercially. The license defines that as giving third parties access to the model for inference or post-training in a way that lets them meaningfully control the inputs, parameters or training data. End-user products that embed the model in specific features, simple relaying of requests to models hosted elsewhere, and internal use are not covered.[17] Mistral said a full technical report was "coming soon".[18]

Mistral published the weights in two official repositories. The main repository stores them in FP8 with 128 x 128 block scaling, split into 54 safetensors files; its configuration leaves the vision tower, the multimodal projector, the output head and some attention projections unquantized.[23] A second repository, Mistral-Large-4-1T-A52B-NVFP4, uses NVFP4 for the mixture-of-experts layers and FP8 for the other linear layers. Mistral says this version targets [NVIDIA Blackwell](https://aiwiki.ai/wiki/nvidia_blackwell) GPUs (B200, B300, GB200 and GB300) and matches the FP8 checkpoint on its evaluations, which covered tests including DeepSWE, SWE-bench Pro, Terminal-Bench 4 and CyberGym.[24] As of October 30, 2026, the Hugging Face API reported 1,239 downloads and 1,270 likes for the FP8 repository and 315 downloads for the NVFP4 repository.[23][24] The model card also links [GGUF](https://aiwiki.ai/wiki/gguf) conversions made by the llama.cpp project and by Unsloth.[17]

Besides the Mistral API, Mistral Studio and OpenRouter, Mistral listed NVIDIA-hosted endpoints and an [NVIDIA NIM](https://aiwiki.ai/wiki/nvidia_nim) container for serving with vLLM, and said the model is available through Microsoft Foundry with private deployment options for Azure developers. In a statement included in Mistral's post, Microsoft's Tina Schuchman described Microsoft as "the first cloud provider to bring Mistral Large 4 to customers".[18] Pierre Stock, Mistral's VP of Science, said at a press briefing reported by The New Stack that smaller versions would follow: "There's going to be a family of models, and one of them is going to be on a laptop."[29]

Alongside the release, Mistral said that while monitoring cybersecurity-related use of the preview in live traffic, it had seen the model refuse black-box attacks against live targets but saw "a potential for uplift to conduct real attacks". It said it was modifying its guardrails and would keep updating them in the coming months. It is also giving vetted enterprise design partners, such as Orange Cyberdefense, preview access to a specialized cybersecurity variant for penetration testing, vulnerability assessment and red-teaming on systems they own or are authorized to test.[18]

Public preview is a distinct stage in Mistral's model lifecycle. Mistral permits silent updates during this stage and does not guarantee that a preview will reach general availability. Results obtained from the preview can therefore describe a changing service rather than a permanently fixed checkpoint.[6] On October 30, 2026, after the weights release, Mistral's API model card still gave the model's status as Public Preview.[3]

## Architecture and specifications

The specifications below come from Mistral's API model card and the Hugging Face model card for the released weights.[3][17]

| Property | Published value |
|---|---|
| Architecture | Granular [mixture of experts](https://aiwiki.ai/wiki/mixture_of_experts)[3] |
| Total parameters | 1.05 trillion in the [model card](https://aiwiki.ai/wiki/model_card)[3] |
| [Vision encoder](https://aiwiki.ai/wiki/vision_encoder) | 1.6 billion parameters[3]; a [Pixtral](https://aiwiki.ai/wiki/pixtral) encoder with 30 layers, a hidden size of 2,048, 16 attention heads and 14 x 14 pixel patches[26] |
| Active parameters | 52 billion in the [model card](https://aiwiki.ai/wiki/model_card) and launch announcement[3][2] |
| Experts | 128 routed experts with 4 active per token, plus 1 shared expert[17] |
| Layers | 72 decoder layers; the first uses a dense feed-forward network and the other 71 use MoE blocks[23][26] |
| Hidden size and attention heads | 6,144; 64 attention heads[23][26] |
| Attention | [Multi-head latent attention](https://aiwiki.ai/wiki/multi-head_latent_attention) with a [DeepSeek Sparse Attention](https://aiwiki.ai/wiki/deepseek_sparse_attention) indexer that selects up to 2,048 key positions per query; key-value state is shared across some layers[26] |
| Position encoding | [YaRN](https://aiwiki.ai/wiki/yarn)-scaled RoPE with one learned multiplier per frequency (Mistral calls it LeRope); three RoPE layers for every layer without positional encoding[26] |
| Vocabulary | 204,800 tokens[23] |
| Weight formats | FP8 (main repository) and NVFP4[23][24] |
| [Speculative decoding](https://aiwiki.ai/wiki/speculative_decoding) | Built-in DSpark draft head: 3 dense MLA draft layers in BF16 that propose 3 tokens at a time[17] |
| Reasoning effort | Low, medium or high (default)[17] |
| Weights license | Mistral Ecosystem Open-Weight License (modified MIT)[17] |
| Mistral [context window](https://aiwiki.ai/wiki/context_window) | 1 million tokens in the [model card](https://aiwiki.ai/wiki/model_card)[3] |
| [OpenRouter](https://aiwiki.ai/wiki/openrouter) context limit | 1,048,576 tokens; one of its Mistral endpoints lists 524,288[4] |
| [OpenRouter](https://aiwiki.ai/wiki/openrouter) maximum output | 256K tokens[4] |
| Input and output | Text and image input; text output on [OpenRouter](https://aiwiki.ai/wiki/openrouter)[4] |

The Hugging Face model card explains the two active-parameter counts: 49B per token, or 52B including embeddings and [output layer](https://aiwiki.ai/wiki/output_layer)s.[17] The model card and current launch article both list 52B.[3][2]

Hugging Face's launch post describes two separate forms of sparsity: the router picks 4 of the 128 experts for each token (5 active experts counting the shared one), while the attention indexer picks which earlier positions each query attends to.[26]

One of OpenRouter's Mistral endpoints lists a lower 524,288-token limit, so the 1-million-token specification should not be assumed for every serving route.[4]

Mistral reports training on 3,800 NVIDIA Grace Blackwell GPUs in its European [data center](https://aiwiki.ai/wiki/data_center)s, which serve the preview, and training data spanning more than 160 languages.[2]

## Self-hosting and customization

Mistral's model card names [vLLM](https://aiwiki.ai/wiki/vllm) as the recommended inference engine and gives instructions for [SGLang](https://aiwiki.ai/wiki/sglang), [Hugging Face Transformers](https://aiwiki.ai/wiki/transformers_library) and [llama.cpp](https://aiwiki.ai/wiki/llama_cpp); for fine-tuning it points to [Axolotl](https://aiwiki.ai/wiki/axolotl), [Unsloth](https://aiwiki.ai/wiki/unsloth) and TRL. Its example vLLM and SGLang commands use tensor parallelism across eight GPUs, and it recommends a temperature of 1.0 and a top_p of 0.95 at every reasoning effort.[17] Hugging Face notes that reasoning cannot be turned off: a request with reasoning_effort set to "none" still runs at high.[26]

The DSpark draft head ships inside the checkpoint. Mistral says it speeds up decoding without changing the target model's outputs and measured up to 1.9 times the throughput of serving the target model alone, with actual gains depending on hardware and workload.[17]

vLLM's deployment recipe for the model listed these checkpoints and validated hardware on October 30, 2026:[25]

| Checkpoint | Weights | Approximate size | Validated on |
|---|---|---:|---|
| Mistral-Large-4-1T-A52B (FP8) | FP8, block-scaled | 1.06 TB | GB300, H200, AMD Instinct MI350X |
| Mistral-Large-4-1T-A52B-NVFP4 | NVFP4 experts; FP8 attention and shared experts | 610 GB | GB300, B200 |
| amd/Mistral-Large-4-1T-A52B-Quark-MXFP4 | MXFP4 routed experts, rest unchanged | 578 GB | AMD MI350 Series |

According to the recipe, the FP8 checkpoint needs about 128 GiB of weights per GPU when spread over eight GPUs and the NVFP4 checkpoint about 76 GiB. Configurations it validated include one four-GPU GB300 tray with NVFP4, eight B300 GPUs, eight [B200](https://aiwiki.ai/wiki/nvidia_b200) GPUs with NVFP4, and two eight-GPU [H200](https://aiwiki.ai/wiki/nvidia_h200) nodes with FP8. The checkpoint advertises 1,048,576 positions, but vLLM validated contexts up to 256K, and at that date no vLLM release or nightly build included support; the recipe used a dedicated Docker image.[25]

Hugging Face offers a ready-made Inference Endpoints configuration on eight H200 GPUs serving up to 131K tokens of context. It reported about 65 tokens per second for a single request on that setup. The llama.cpp team's 2-bit quantizations run on a single Apple Silicon Ultra machine with 512 GB of unified memory; on an M3 Ultra, Hugging Face measured about 90 tokens per second of prefill and 15 tokens per second of generation with the Q2_K file.[26]

For managed customization Mistral points enterprises to [Mistral Forge](https://aiwiki.ai/wiki/mistral_forge). When it announced the preview, Mistral said the model was trained with "the same training, customization, and RL environment we offer our customers through Mistral Forge".[2] At the open-weight release it said customers were tailoring the model "with proprietary security data, policies, detection logic, and threat environments" through Forge, and that for well-defined cybersecurity tasks the model can be optimized into smaller task-specific variants.[18]

On October 29, 2026, a vLLM blog post credited to "Mistral and the vLLM Team" reported day-0 reinforcement learning support for the model in [SkyRL](https://aiwiki.ai/wiki/skyrl), an open-source RL library. The trainer ran rank-32 [LoRA](https://aiwiki.ai/wiki/lora) on top of the NVFP4 checkpoint, with vLLM generating rollouts and either FSDP2 or Megatron training, on eight GB200 nodes. On the [DAPO](https://aiwiki.ai/wiki/dapo) recipe, the post said, [AIME](https://aiwiki.ai/wiki/aime) 2024 accuracy averaged over eight samples (avg@8) rose from about 64% to about 75% within 75 steps.[31] SkyRL's tracking issue shows that the run continued to 105 steps, with accuracy at 0.72-0.73 at steps 95-100. The code was in five stacked draft pull requests (#2576 to #2580) submitted from a contributor's fork, which were still open and unmerged on October 31, 2026.[32]

## API access and pricing

Mistral's Chat Completions example uses `mistral-large-4` with `reasoning_effort="high"`.[3]

[OpenRouter](https://aiwiki.ai/wiki/openrouter) identifies its route as `mistralai/mistral-large-4-0`.[4] These are provider-specific identifiers, not interchangeable request strings.

Mistral cut the price when it released the weights: "Starting today, Mistral Large 4 is priced at $0.70 input / M, $0.07 cached input / M, and $2.10 output / M tokens."[18] Its pricing page and model card showed the same standard-tier rates on October 30, 2026, with no sale label.[5][3]

The preview launched at a higher list price with a 50% discount for its first two weeks, which Mistral's changelog announced.[1] Archived copies of the model card show the list and sale prices on October 12 and the full list price again on October 28, after the sale had ended.[19]

| Token type (US dollars per million tokens) | Preview list price (October 6) | Two-week launch sale | Price from October 29, 2026 |
|---|---:|---:|---:|
| Input | $1.36[19] | $0.68[19] | $0.70[18] |
| Cached input | $0.14[19] | $0.07[19] | $0.07[18] |
| Output | $4.18[19] | $2.09[19] | $2.10[18] |

Mistral's pricing page separates standard, batch, and priority tiers.[5]

## Tool use and document workflows

The card lists chat completions, [structured output](https://aiwiki.ai/wiki/structured_output)s, function calling, Document QnA, prefix completion, batching, Agents and Conversations, and built-in tools.[3]

[Function calling](https://aiwiki.ai/wiki/function_calling) lets an application provide tool definitions and receive a selected function with generated arguments. In the ordinary Chat Completions workflow, the application executes that function and supplies the result back to the model. Mistral separately offers server-side tools through its Agents and Conversations APIs. Tool-calling support does not mean that an ordinary completion has unrestricted access to an application's systems.[7]

For structured output, Mistral provides JSON mode and custom schemas. Its documentation recommends custom structured outputs when a specific format is required; JSON mode still requires the prompt to request JSON and describe the intended format.[8]

Document QnA combines an OCR stage, which extracts document text and structure, with a [language model](https://aiwiki.ai/wiki/language_model) that answers questions about the extracted content. The documented workflow covers information extraction, summarization, and comparisons across documents.[9]

Agents store a model selection, instructions, tools, and completion settings. Conversations record interactions and can be started with either an agent or a model directly. Mistral's built-in tool types include web search, a code interpreter, image generation, and a document library.[10] In particular, a tool for image generation is a platform capability, not evidence that the model's own output modality includes generated images.

## Evaluations

### Results reported by Mistral

The October 6 launch announcement reported these preview results.[2]

| Evaluation | Reported result |
|---|---:|
| DeepSWE v1.1 | 61.7%[2] |
| Terminal-Bench 4 | 28.3%[2] |
| Cybench | 93%[2] |
| Lakera B3 attack resistance | 93.3%[2] |

The original [Cybench](https://aiwiki.ai/wiki/cybench) research describes 40 capture-the-flag tasks from four competitions, with command-execution environments and optional subtask guidance. It also studies different agent scaffolds.[15] Its original experiments predate Mistral Large 4; the paper defines the benchmark rather than independently validating this model's launch score.

For the open-weight release on October 29, Mistral published these results, several of them measured by outside evaluators.[18]

| Evaluation | Reported result |
|---|---:|
| Cognition FrontierCode 2.0 (private benchmark) | 19.7%[18] |
| Artificial Analysis Coding Agent Index | 48%[18] |
| Terminal-Bench 4.0 (Artificial Analysis Coding Agent Index component) | 41%, which Mistral compares with the preview's self-reported 28% as a 13-point gain, with 30% fewer output tokens[18] |
| Artificial Analysis Cyber Index | 51%[18] |
| FinWorkBench (Finch) | 73.3%[18] |
| BrowseComp | 84.5%[18] |
| Refusal of malicious cyber requests (cyber prompts from JailbreakBench, StrongREJECT and AgentHarm) | 96.6% average[18] |
| B3 Agent Security Benchmark attack resistance | 97.1%, against 93.3% for the preview[18] |

Mistral describes the Coding Agent Index score as "effectively tied" with GLM-5.3, the leading open-weight model on that index, while using 47% fewer tokens per task on average.[18] In blind side-by-side coding comparisons run by Surge, graders preferred Mistral Large 4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, but preferred Claude Opus 5.5 over it.[18]

### FrontierCode 2.0

Mistral's FrontierCode figure, and its claim on X that the model "ranks #1 on @cognition's FrontierCode among open models", refer to [Cognition](https://aiwiki.ai/wiki/cognition_ai)'s public leaderboard.[30][27] [FrontierCode](https://aiwiki.ai/wiki/frontiercode) 2.0 became the leaderboard's current version on October 27, 2026. It has 132 tasks built with open-source maintainers and grades whether a maintainer would merge the agent's change. The same day, Cognition moved Mistral Large 4 to the Mistral Vibe harness at low, medium and high reasoning, "replacing the Devin CLI preview results".[27] Its data file on October 30 gave these best scores:[27]

| Model (reasoning effort) | Harness | FrontierCode 2.0 score | Mean cost per rollout |
|---|---|---:|---:|
| [Claude Opus 5.5](https://aiwiki.ai/wiki/claude_opus_5_5) (max) | Claude Code | 45.5% | $12.82 |
| Mistral Large 4 (high) | Mistral Vibe | 19.7% | $1.55 |
| [DeepSeek V4.1 Flash](https://aiwiki.ai/wiki/deepseek_v4_1_flash) (max) | Devin CLI | 19.3% | $0.38 |
| [GLM-5.3](https://aiwiki.ai/wiki/glm_5_3) (max) | Devin CLI | 17.8% | $7.52 |
| [Kimi K3](https://aiwiki.ai/wiki/kimi_k3) (high) | Devin CLI | 17.7% | $4.48 |

Mistral Large 4 led the other open-weight models listed by less than half a point, while 12 of the 19 models shown on the leaderboard had higher best scores (the data file also holds results for seven superseded models that the page shows only as lineage history). Its low and medium settings scored 9.5% and 8.8%. The comparison mixes harnesses: Mistral's model ran in Mistral's own agent, while the other open-weight models ran in Cognition's Devin CLI.[27] Artificial Analysis, which runs FrontierCode 2.0 in its own Stirrup harness for its Intelligence Index, measured Mistral Large 4 (High) at 15.0%, below open-weight models such as DeepSeek V4.1 Flash 1017 (28.5%), MiMo-V2.6-Pro (24.0%) and GLM-5.3 (21.6%) in the same runs.[20][22]

### Independent evaluator snapshots

[Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) listed Mistral Large 4 Preview with an Intelligence Index score of 38 on October 7, 2026. Its page identified the index version as 4.3.2, combining ten evaluations, and reported 116.1 output [tokens per second](https://aiwiki.ai/wiki/tokens_per_second) for Mistral's API. At that time it labeled the preview proprietary because the weights were not yet public.[11]

On October 30, 2026, after the weights release, Artificial Analysis gave Mistral Large 4 (High) a score of 26 on version 5.0 of its Intelligence Index. Version 5.0 changed the evaluation set: it added HLE-Diamond, Terminal-Bench-Science and FrontierCode 2.0 and removed GDPval-AA, SciCode, AA-LCR and Humanity's Last Exam, so the score of 26 is not on the same basis as the earlier 38.[20][22] The harness changes Terminal-Bench 4.0 results considerably: Artificial Analysis measured 24.7% for Mistral Large 4 (High) with the mini-swe-agent harness it uses for Terminal-Bench 4.0 in the Intelligence Index,[20][22] and 41.4% when the model ran inside Mistral's own Vibe coding agent for its Coding Agent Index.[21] Mistral's reported 41% is the second run. The Artificial Analysis page also listed 161 output tokens per second and an average cost of $2.06 per Intelligence Index task on October 30.[20]

Artificial Analysis's Coding Agent Index averages FrontierCode v1.1, SWE-Atlas-QnA and Terminal-Bench 4.0. Its data on October 30 gave Mistral Vibe with Mistral Large 4 (High) 48.3%, against 48.9% for GLM-5.3 (Max) running in OpenCode. The Mistral run averaged about 8.5 million total tokens and $1.26 per task, against about 15.9 million tokens and $5.23 for the GLM-5.3 run, a token reduction of about 47% that matches Mistral's figure.[21] Mistral's X post described the result as "a new benchmark for Pareto efficiency".[30]

On the Artificial Analysis Cyber Index, which combines CWE-Bench-AA, DeepsecBench-AA and CyberGym-E2E-AA, Mistral Large 4 (High) scored 50.9 on October 30. That was seventh among the models listed and third among the models Artificial Analysis classed as open-weight (Moonshot's Kimi K3.1 Pro, at 64.1, was counted as proprietary because its promised weights were not yet out), behind DeepSeek V4.1 Flash 1017 (Max) at 56.8 and MiMo-V2.6-Pro at 56.1. The released model scored 79.4% on CyberGym-E2E-AA, compared with 81.7% for the preview.[28]

On October 7, 2026, while the model was in preview, [Vals AI](https://aiwiki.ai/wiki/vals_ai) published the following results for Mistral Large 4. The displayed uncertainty values are retained with the scores.[12] Vals has since replaced them with results on a different benchmark set, so these figures describe that earlier snapshot rather than the released checkpoint.

| Vals evaluation | Score |
|---|---:|
| Vals Index | 48.05% +/- 1.11[12] |
| Finance Agent v2 | 54.68% +/- 0.58[12] |
| Harvey's Legal Agent Benchmark | 15.83% +/- 2.96[12] |
| Terminal-Bench 4.0 | 22.73% +/- 0.88[12] |

Vals' methodology uses controlled evaluation harnesses, private [test set](https://aiwiki.ai/wiki/test_set)s for proprietary benchmarks, and published standard errors. For benchmarks with multiple runs, including Terminal-Bench and Finance Agent v2, it calculates uncertainty over run-level scores. Such error bars do not capture every change in prompts or deployment settings.[13] Its Terminal-Bench result is separate from Mistral's reported 28.3% for the preview; the two figures should not be collapsed into one score.

Harvey's Legal Agent Benchmark evaluates legal deliverables using file-system tools and skills for documents, presentations, and spreadsheets. Vals reports results on a held-out task set and distinguishes whole-task success from satisfying individual criteria.[14] A task pass rate on this benchmark is not a general measure of legal accuracy or an authorization to rely on generated legal advice.

## Partner statements

Mistral's October 29 post included statements from cybersecurity vendors and agencies, several of which said they had tested or evaluated the model; its announcement on X named CrowdStrike, [Cloudflare](https://aiwiki.ai/wiki/cloudflare), Rubrik and Zscaler "and others".[18][30] The statements are testimonials published by Mistral. None gives numerical test results.

| Organization | Speaker | Statement (excerpt) |
|---|---|---|
| CrowdStrike | Bartley Richardson, Chief AI and Autonomous Systems Officer | "Our work with Mistral Large 4 tests how well frontier AI can reason through complex vulnerabilities, accelerate remediation, and ensure defenders move faster than adversaries." |
| Cloudflare | Stephanie Cohen, Chief Strategy Officer | "Mistral's ML4 lets defenders audit the architecture, deploy on their own infrastructure, and keep full sovereignty." |
| Rubrik | Anneka Gupta, Chief Product Officer | "we'll continue rigorous cybersecurity testing with Mistral Large 4 and work closely together to serve evolving regional demands." |
| Zscaler | Dhawal Sharma, EVP of AI and Data Security | "Zscaler Security for AI offerings, delivered on our sovereign Zero Trust Exchange in Europe, now integrate with Mistral" |
| Booz Allen Hamilton | Brad Medairy, President, National Cyber | "The model was able to match vulnerability detection recall rates against other frontier cyber models with a compelling cost curve." |
| Cisco | Amy Chang, Head of AI Threat Intelligence and Security Research | Open-weight cybersecurity models "allow researchers to study how these models behave, including where they could be misused." |
| Elastic | James Spiteri, Senior Director, Product Management | "We've already evaluated ML4 against real cybersecurity workloads" |
| ENISA | Laura Heuvinck, Head of Communications | The agency is testing the model "to identify security vulnerabilities, to audit code and analyse defensive logs." |
| Fortinet | John Whittle, Chief Operating Officer | "Innovations like Mistral Large 4 represent promising steps forward in AI's potential to transform cyber defense." |
| Orange Cyberdefense | Hugues Foulon, CEO | "With Mistral AI, we are building European cyber AI, trained by our experts and operated under their control." |

Other statements in the post came from NVIDIA, Microsoft, Hugging Face and ZML, and from customers including La Banque Postale, ABN AMRO, Belfius, Crédit Agricole and the European Space Agency's Φ-lab.[18]

## Safety evaluation scope

Mistral's launch report discusses resistance to indirect [prompt injection](https://aiwiki.ai/wiki/prompt_injection) and refusals of malicious cybersecurity requests, including tests drawn from [JailbreakBench](https://aiwiki.ai/wiki/jailbreakbench), [StrongREJECT](https://aiwiki.ai/wiki/strongreject), and AgentHarm.[2] These evaluations address specified attack or misuse settings.

[AgentHarm](https://aiwiki.ai/wiki/agentharm) contains 110 malicious agent tasks, expanded to 440 through augmentations, across 11 harm categories. Its research evaluates both refusals and whether an attacked agent can complete a harmful multistep workflow.[16] A refusal measurement and an attack-resistance measurement test different behavior; neither constitutes a guarantee that an agent is safe in every deployment.

## References

1. Mistral AI. ["Changelog," October 6, 2026 entry](https://docs.mistral.ai/resources/changelogs). Accessed October 12, 2026.
2. Mistral AI. ["Introducing Mistral Large 4"](https://mistral.ai/news/mistral-large-4/). October 6, 2026. Accessed October 12, 2026.
3. Mistral AI. ["Mistral Large 4" model card, including Weights and Usage tabs](https://docs.mistral.ai/models/mistral-large-4-0). October 6, 2026. Accessed October 30, 2026.
4. OpenRouter. ["Mistral: Mistral Large 4"](https://openrouter.ai/mistralai/mistral-large-4-0). Accessed October 30, 2026.
5. Mistral AI. ["Pricing"](https://docs.mistral.ai/inference/pricing). Accessed October 30, 2026.
6. Mistral AI. ["Model lifecycle policy"](https://docs.mistral.ai/inference/model-lifecycle). Accessed October 7, 2026.
7. Mistral AI. ["Function Calling"](https://docs.mistral.ai/studio/conversations/function-calling). Accessed October 7, 2026.
8. Mistral AI. ["Structured Outputs"](https://docs.mistral.ai/studio/conversations/structured-output). Accessed October 7, 2026.
9. Mistral AI. ["Document AI QnA"](https://docs.mistral.ai/studio/document-processing/document_qna). Accessed October 7, 2026.
10. Mistral AI. ["Agents & Conversations"](https://docs.mistral.ai/studio/agents/agents-api). Accessed October 7, 2026.
11. Artificial Analysis. ["Mistral Large 4 Preview: Intelligence, Performance & Price Analysis"](https://artificialanalysis.ai/models/mistral-large-4). Accessed October 7, 2026.
12. Vals AI. ["Mistral Large 4 Benchmarks, Cost and Capabilities"](https://www.vals.ai/models/mistralai_mistral-large-4). Accessed October 7, 2026 (figures cited) and October 30, 2026 (updated page); October 7 version archived at [Wayback Machine](https://web.archive.org/web/20261007110318/https://www.vals.ai/models/mistralai_mistral-large-4).
13. Vals AI. ["Methodology"](https://www.vals.ai/methodology). Accessed October 7, 2026.
14. Vals AI. ["Harvey's Legal Agent Benchmark"](https://www.vals.ai/benchmarks/hlab). Updated October 6, 2026. Accessed October 7, 2026.
15. Zhang, Andy K., et al. ["Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models"](https://arxiv.org/abs/2408.08926). arXiv:2408.08926, submitted August 15, 2024; revised April 12, 2025. ICLR 2025.
16. Andriushchenko, Maksym, et al. ["AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents"](https://arxiv.org/abs/2410.09024). arXiv:2410.09024, submitted October 11, 2024; revised April 18, 2025. ICLR 2025.
17. Mistral AI. ["Mistral-Large-4-1T-A52B" model card and license](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B). Hugging Face. Updated October 29, 2026. Accessed October 30, 2026.
18. Mistral AI. ["Mistral Large 4 (Le Chonk) is now open weight"](https://mistral.ai/news/mistral-large-4-oss/). October 29, 2026. Accessed October 30, 2026.
19. Internet Archive. Wayback Machine captures of Mistral's "Mistral Large 4" model card, [October 12, 2026](https://web.archive.org/web/20261012201412/https://docs.mistral.ai/models/mistral-large-4-0) and [October 28, 2026](https://web.archive.org/web/20261028150002/https://docs.mistral.ai/models/mistral-large-4-0). Accessed October 30, 2026.
20. Artificial Analysis. ["Mistral Large 4: Intelligence, Performance & Price Analysis"](https://artificialanalysis.ai/models/mistral-large-4). Accessed October 30, 2026.
21. Artificial Analysis. ["Coding Agent Index"](https://artificialanalysis.ai/agents/coding-agents). Accessed October 30, 2026.
22. Artificial Analysis. ["Artificial Analysis Intelligence Benchmarking Methodology"](https://artificialanalysis.ai/methodology/intelligence-benchmarking), including the version history. Accessed October 30, 2026.
23. Mistral AI. ["Mistral-Large-4-1T-A52B" configuration files (config.json, params.json)](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/config.json) and [Hugging Face Hub API record](https://huggingface.co/api/models/mistralai/Mistral-Large-4-1T-A52B). Accessed October 30, 2026.
24. Mistral AI. ["Mistral-Large-4-1T-A52B-NVFP4" model card](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B-NVFP4) and [Hugging Face Hub API record](https://huggingface.co/api/models/mistralai/Mistral-Large-4-1T-A52B-NVFP4). Hugging Face. Accessed October 30, 2026.
25. vLLM project. ["mistralai/Mistral-Large-4-1T-A52B" recipe](https://recipes.vllm.ai/mistralai/Mistral-Large-4-1T-A52B). vLLM Recipes. Accessed October 30, 2026.
26. Noyan, Merve, Alvaro Bartolome, et al. (Hugging Face). ["Welcome Le Chonk: Mistral Large 4 on Hugging Face"](https://huggingface.co/blog/mistral-large4). October 29, 2026. Accessed October 30, 2026.
27. Cognition. ["FrontierCode Leaderboard"](https://cognition.ai/frontiercode), including its changelog, and the leaderboard [data file](https://cognition.ai/data/frontier-code-2.0/data.json). Accessed October 30, 2026.
28. Artificial Analysis. ["Artificial Analysis Cyber Index"](https://artificialanalysis.ai/evaluations/artificial-analysis-cyber-index). Accessed October 30, 2026.
29. Shubel, Meredith. ["Mistral's Large 4 open weights arrive with a 1-point coding lead over GLM-5.3: 'There's nothing you can't do with this model that you can do with competitors.'"](https://thenewstack.io/mistral-chonk-coding-model/). The New Stack, October 29, 2026. Accessed October 30, 2026.
30. Mistral AI (@MistralAI). [Post announcing the open-weight release](https://x.com/MistralAI/status/2115811242556137820). X, October 29, 2026. Accessed October 30, 2026.
31. Mistral and the vLLM Team. ["Day 0 support for Mistral Large 4 on vLLM"](https://vllm.ai/blog/2026-10-29-mistral-large-4), section "Day 0 Support for RL with SkyRL". vLLM Blog, October 29, 2026. Accessed October 31, 2026.
32. NovaSky-AI/SkyRL issue #2575. ["[RFC] Day-0 RL support for Mistral Large 4 (NVFP4) on FSDP and Megatron"](https://github.com/NovaSky-AI/SkyRL/issues/2575), and draft pull requests [#2576](https://github.com/NovaSky-AI/SkyRL/pull/2576) to [#2580](https://github.com/NovaSky-AI/SkyRL/pull/2580). GitHub, opened October 29, 2026. Accessed October 31, 2026.

