Mistral Large 4 Goes Open-Weight: What the Independent Scores Support

Mistral published the downloadable weights of Mistral Large 4 on 29 October 2026 and says it leads open-weight models in defensive security, but Artificial Analysis's Cyber Index scores it 51, third among open-weight models behind 57 and 56.⁠¹⁠² The case for running it on your own infrastructure is that access cannot be withdrawn, and that case holds, though the published serving recipes call for four to eight current-generation data-center GPUs.⁠³⁠⁴ Confidence: moderate on the benchmark gap and the hardware, low on dual-use risk, because the only misuse assessment is Mistral's own.

What Mistral released on 29 October

Mistral announced Mistral Large 4 on 6 October as an API preview and said the weights would follow by the end of the month.⁠⁵ On 29 October it published them on Hugging Face after what it describes as three more weeks of reinforcement learning, so the downloadable model differs from the one previewed.⁠¹ The model is a mixture-of-experts system with about one trillion parameters, and only a subset of those runs for each token.

Sources disagree on how many parameters are active per token. The Register put it at about 49 billion, SGLang's documentation says 52 billion, and NVIDIA's Dynamo recipe says 54 billion.⁠⁵⁠⁴⁠⁶ LLM Rumors reports that Mistral's model card uses 49 billion when embeddings and output layers are excluded and 52 billion when they are counted, and gives no reason for NVIDIA's figure.⁠⁷ Storage needs follow the total count, so the disagreement matters less for hardware planning than it first appears.

The weights ship under the Mistral Ecosystem Open-Weight License, a modified MIT license that allows use, modification, fine-tuning and commercial use. One condition applies: businesses that operate a Model-as-a-Service offering need a separate commercial license from Mistral.⁠¹ Mistral's API costs $0.70 per million input tokens and $2.10 per million output tokens, and Microsoft Foundry also offers the model as a managed service.⁠¹

What was measured, by whom, and who paid

Almost every capability number in the coverage was produced or selected by Mistral. Its launch post carries charts from Artificial Analysis, Cognition and Surge AI evaluations alongside Mistral's own refusal and agent-security tests.⁠¹ Artificial Analysis is an independent benchmarking firm, but this article read its Cyber Index and Coding Agent Index results through Implicator and not on Artificial Analysis's own pages.⁠² The Cyber Index, as Mistral describes it, is a security benchmark of vulnerability detection, validation and remediation: it measures how well a model finds vulnerabilities in a codebase, reproduces and validates them, and patches them without breaking existing behavior.⁠¹

Only the SGLang figures come with a full stated protocol, and they measure serving speed, not capability.

Where Mistral's claims and the independent scores disagree

On coding the gap is small. Reported Mistral says its 48% on the Artificial Analysis Coding Agent Index is effectively tied with GLM-5.3, the leading open-weight model, while using 47% fewer tokens per task.⁠¹ Implicator reports that Artificial Analysis ranks Mistral Large 4 second, one point behind GLM-5.3, and that a one-point difference alone does not establish statistical significance.⁠² The two accounts agree on the figures and differ on the label. The independent benchmark supports the efficiency claim and not the leadership claim.

On Terminal-Bench 4.0, a test of coding agents, Mistral reports 41%, up 13 points from its preview and the lead among open-weight models.⁠¹ Artificial Analysis's Intelligence Index assessment reports 25% on the same test, which Implicator attributes to different reported evaluation contexts.⁠² No source here replicates either number, so benchmark reproducibility is untested for both.

Measured On the Cyber Index, Artificial Analysis confirms the 51 that Mistral reports and ranks it third among open-weight models, behind DeepSeek V4.1 Flash 1017 at 57 and MiMo-V2.6-Pro at 56.⁠² Mistral's launch post still says the model leads open-weight benchmarks in defensive security,⁠¹ and Stock told Fortune it is the best open-weight model on cyber and coding.⁠⁸ The independent score is the more credible, because a third party scored it and the leadership claim comes from the seller. Mistral does lead on the broader Intelligence Index among open-weight models developed outside China, at 26, one point above Meta's Muse Spark Open, against 32 for DeepSeek V4.1 Flash 1017 and GLM-5.3.⁠²

The coding claim is close to the independent score, and the cyber claim is not.

GetCoreTech's earlier analysis of Gemini 4 Carbon's cybersecurity benchmarks covers a different vendor's cyber results, and this article makes no cross-model comparison with it.

Why organizations choose their own infrastructure, and what it takes

Mistral's argument for self-hosted AI is about control. Stock told TechCrunch that organizations can use the model to defend against cyberattacks with access that cannot be cut off, and that it can be deployed on-premises in a company's own environment.⁠³ Mistral also argues that closed providers' refusals can block legitimate vulnerability research and incident response, a problem The Register has covered separately.⁠⁵ Customers can tailor the model with proprietary data through Mistral's Forge product, and The Register reports it was trained on about 3,800 Grace Blackwell GPUs in Mistral's own European data centers, which Mistral presents as clear provenance for sovereignty requirements.⁠¹⁠⁵

Partner testimonials in the launch post, from firms including Cloudflare, Cisco, CrowdStrike and Booz Allen, were selected by Mistral and mostly describe testing or intent. Booz Allen said the model matched the vulnerability-detection recall of other frontier cyber models on an internal benchmark at a good cost, and gave no figures.⁠¹ Mistral's chief revenue officer said business results from systems built on the new model are not yet available.⁠²

The weights are large. LLM Rumors, citing the vLLM serving recipe, reports about 1.06 TB for the FP8 model checkpoint and 610 GB for the NVFP4 checkpoint, before memory for the key-value cache and concurrent requests.⁠⁷ SGLang's verified recipes use eight B300 GPUs for FP8 and four GB300 GPUs for NVFP4, and its documentation marks other GPUs as not verified.⁠⁴ NVIDIA's Dynamo recipe for the NVFP4 checkpoint deploys eight GB300 GPUs as two four-GPU workers, each requesting 704 GiB of host memory.⁠⁶ The Register estimated that eight-GPU systems such as HGX B300 or AMD's MI355X could run the model, noting that Mistral gave no hardware guidance at the preview.⁠⁵

Measured On SGLang's stated protocol, the FP8 setup on eight B300 GPUs reached an inference throughput of 774 total tokens per second per GPU at 64 concurrent requests, and the NVFP4 setup on four GB300 GPUs reached 1,560.⁠⁴ The model declares a native context window of 1,048,576 tokens, but LLM Rumors reports that the vLLM recipe's validation reached 256,000.⁠⁷ Benchmark costs do not capture self-hosting. Artificial Analysis measured $2.06 per Intelligence Index task for the released model against $5.54 for the preview, DeepSeek and MiMo score higher at lower measured cost, and Implicator notes these figures exclude hosting, customization and human review.⁠²

Dual-use risk: Mistral names it, and no independent source measured it

Reported Mistral's launch post says that after monitoring cybersecurity-related use in its live traffic since the preview, it saw the model refuse black-box attacks against live targets and still sees "a potential for uplift to conduct real attacks," and that it is modifying its AI guardrails.⁠¹ It also gives vetted design partners, including Orange Cyberdefense, preview access to a specialized variant for penetration testing, vulnerability assessment and red-teaming, under contractual safeguards and with testing limited to authorized systems.⁠¹

Mistral reports a 96.6% average refusal rate on malicious cyber requests, the highest it measures among open models, on prompts from JailbreakBench, StrongREJECT and AgentHarm.⁠¹ A refusal rate is the share of harmful requests a model declines. The test is vendor-run, and no source here tests whether that refusal behavior survives fine-tuning by whoever runs the weights.

Opinion Stock's argument for release and the risk share one property. He told TechCrunch that defenders need frontier capability because malicious actors may use technology of their own to attack,⁠³ and said at a press briefing, as Implicator reports, that once weights are public access cannot be cut off.⁠² Mistral's monitoring covered its own live traffic, and a downloaded copy runs outside it, by construction.

Sophos threat-intelligence director Rafe Pilling told Wired the technology is clearly dual-use, as Implicator reports, and said the safeguard is monitoring and controls able to detect deceptive behavior.⁠²

What the evidence does not establish

The cyber results measure AI-assisted vulnerability discovery, reproduction and patching on benchmark tasks, and no source here reports performance inside a buyer's own environment. Mistral's partners describe evaluation and intent, and its chief revenue officer said operational results are not yet available.⁠¹⁠²

No source prices a self-hosted deployment, so the claim that owning the model lowers cost is untested. Artificial Analysis's per-task costs exclude hosting, customization and review.⁠²

General capability against closed models points the other way. The Register reported that Artificial Analysis's early testing placed the preview well behind OpenAI and Anthropic flagship models, between DeepSeek V4.1 Flash and GPT6 Luna on its intelligence leaderboard.⁠⁵

Training details are also missing. Mistral says a full technical report is coming, so the training data, methods and evaluation protocols behind its charts cannot yet be audited.⁠¹

Verdict: a credible self-hosting option with an overstated cyber claim

Opinion Open weights give organizations control that an API cannot, at the price of a multi-GPU cluster, and Mistral's claim to the best open-weight cyber model is not supported by the independent score.

The third-place Cyber Index result and the one-point gap on coding both show a competitive release whose leadership claims run ahead of independent measurements.⁠² Confidence stays moderate on capability and hardware, and low on dual-use risk, because the only assessment of misuse is the vendor's own.⁠¹

A re-run of the Cyber Index that puts Mistral Large 4 above 57, or an independent replication of its 41% Terminal-Bench figure, would move the capability verdict. An independent test of whether refusal behavior survives fine-tuning would move the risk verdict, and Mistral's promised technical report would let outsiders audit both.⁠¹