Mistral Large 4 review: licence, API price and hardware

This review is built from what could be checked on November 3, 2026: Mistral AI's preview and open-weight announcements, the model card, technical report and licence on Hugging Face, Mistral's API documentation, prices and terms; independent results from Artificial Analysis and Vals AI; public tests by Metergraph and Simon Willison; user reports on Hacker News and Hugging Face; reporting by CNBC, The Next Web and DataNorth; and the price pages of six rival models. Benchmark figures are Mistral's own claims unless another source is named.

What Mistral Large 4 is

Mistral Large 4 is the flagship model of Mistral AI, the Paris company founded in April 2023 by Arthur Mensch, Guillaume Lample and Timothée Lacroix (about Mistral). The model is a mixture of experts: about 1 trillion parameters in total, of which 49 billion are active for each token, or 52 billion counting the embedding and output layers. Each token goes through 4 of 128 routed experts plus one shared expert. It reads text and images, writes text, holds 1,048,576 tokens of context and can reason at low, medium or high effort (model card). Mistral's docs give 1.05 trillion parameters and a 1.6-billion-parameter vision encoder (Mistral docs).

Mistral trained it from scratch in its own data centre near Paris. The technical report puts pretraining at 45 days on 3,840 Nvidia GB200 GPUs, about 4.15 million GPU-hours, and the training data covered more than 160 languages (technical report, preview post). Mistral pitches the model on that origin, as a "trusted foundation for Sovereign AI", and on two specialities: agentic coding and defensive cybersecurity (announcement).

It replaces two models. Mistral deprecated Mistral Large 3 and Mistral Medium 3.5 on October 31, 2026, and both retire on April 30, 2027 (models overview). Next to Large 3, from Mistral's docs:

Sources: Mistral Large 3, Mistral Large 4.

What changed in October

Mistral launched a public preview on October 6, 2026, API only, and promised the weights by the end of the month. Until then, it said, cybersecurity firms, vetted partners and state authorities were testing a version with reduced moderation and expanded cyber capabilities (preview post, CNBC). The preview had two reasoning levels, none and high (Simon Willison).

On October 29 Mistral published the weights on Hugging Face, together with an NVFP4 version for Nvidia's Blackwell GPUs that it says matches the FP8 original on its evaluations (commit history, NVFP4 card). The released model had three more weeks of reinforcement learning: Mistral reports Terminal-Bench 4.0 up from 28% to 41% with 30% fewer output tokens, and it added low and medium reasoning modes (announcement, The Next Web). The technical report followed on November 2.

The price changed too. During the preview, eesel.ai and Hacker News users reported a list price of 1.36 and 4.18 USD per million input and output tokens, sold at half price (eesel.ai, Hacker News). With the weights, Mistral set 0.70 and 2.10 USD as its price "starting today" (announcement).

Two things have not changed. Mistral's docs still label the API model Public Preview, version 26.10, on November 3 (Mistral docs), and Mistral says it will release no base model with this version (Hugging Face discussion).

What it costs

Mistral sells Mistral Large 4 by the token, on its API price list on November 3, 2026, in USD per million tokens:

Sources: API pricing, pricing page, regional inference.

The per-token price is low for a model this size; the token count is not. Artificial Analysis needed 110 million output tokens to run its Intelligence Index at high effort, against a median of 100 million for open models of its class, and put the cost at 2.06 USD per task. At medium effort the score drops from 26 to 23, but tokens fall to 73 million and the cost to 1.41 USD a task (Artificial Analysis).

For people rather than developers, Mistral's app was renamed from Le Chat to Vibe in May 2026 (Vibe launch). Its plans on the pricing page: a limited free plan, Pro at 14.99 USD a month, Team at 24.99 USD per user a month and custom Enterprise deals. The pricing page promises Mistral's "most powerful" models without naming one per plan; The Next Web reported Large 4 was available to some Vibe users at launch, and Mistral points coders to Vibe Code (The Next Web, announcement).

The licence: free unless you sell it as a service

The weights come under the Mistral Ecosystem Open-Weight License, which Mistral calls a modified MIT licence (licence text). It grants the right to use, copy, modify, fine-tune, distribute and sell the model, free of charge, with one condition: a company that operates a model-as-a-service business must sign a separate commercial licence with Mistral before using the model commercially. The licence defines that as giving third parties access to the model for inference or post-training with meaningful control over inputs, parameters or training data. It does not cover:

• internal use, where the model is not made available to anyone outside the company;
• end-user products that only embed the model's abilities in specific features;
• relaying requests to a model hosted by someone else;
• use through Mistral's own products.

For a bank, ministry or manufacturer running it for its own staff, there is no licence fee. A hosting company that wants to resell it needs a contract, and DataNorth points out that a managed setup for a single client sits in a grey zone the announcement does not settle (DataNorth). Mistral Large 3 was under Apache 2.0, so this is a tightening. On Hacker News, several users argued it should be called "weights available" rather than open; others compared it to the source-available licences of database companies (Hacker News).

Running the weights yourself

Mixture-of-experts makes each token cheap to compute, but every expert still has to sit in memory. The checkpoints and the hardware their makers name:

Sources: FP8 files, NVFP4 files, SGLang cookbook, llama.cpp GGUFs, Unsloth, MLX build.

Hopper GPUs are a poor fit. SGLang's cookbook does not offer one eight-GPU H200 node, because the FP8 weights take about 129 GiB of each 140 GiB card and leave almost nothing for the cache; its two-node H200 command is marked not yet verified. A Hacker News user with four H200 cards, about 540 GB in all, was told by another that even the 610 GB NVFP4 version would not fit (Hacker News). Mistral's technical report describes serving on four GB300 GPUs with about 20 million tokens of cache space (technical report), and its model card recommends Mistral's API "if local serving is subpar" (model card).

Two serving details matter. Mistral's model card says its tool-call and reasoning parsers are required for tool use and to separate reasoning from the answer, and SGLang warns that without them both come back as raw text. And the chat template enforces Mistral's message order: one user found that Claude Code behind a LiteLLM gateway failed in four of four sessions, and Mistral's engineer replied that the order will not be loosened and suggested Vibe or a proxy (Hugging Face discussion).

What the benchmarks and public tests show

Mistral's own numbers. Mistral reports 19.7% on Cognition's FrontierCode 2.0, ahead of GLM-5.3 (17.8%) and DeepSeek V4.1 Flash (19.3%); 41% on Terminal-Bench 4.0; 48% on Artificial Analysis's Coding Agent Index, level with GLM-5.3 at 47% fewer tokens per task; and 51% on the Artificial Analysis Cyber Index, below DeepSeek V4.1 Flash's 57% on the same chart. In blind code comparisons graded by Surge AI, reviewers preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, but preferred Claude Opus 5.5's answers 78% of the time (announcement). The technical report's own tables are less flattering in places: 26.0% accuracy on Artificial Analysis's Omniscience knowledge questions, the lowest of the five models scored, 17.3% on HLE-Diamond against 26.9% for Kimi K3, and 65.2% on DeepSWE against 74.2% for DeepSeek V4.1 Flash (technical report).

Artificial Analysis. Its Intelligence Index v5.0 gives Mistral Large 4 26 at high effort, 10th of 23 open-weight models in its size class, and the model card calls that the highest score among Western open-weight models (Artificial Analysis, model card). Reviews of the preview quote 38; that was the older v4 index (Simon Willison, eesel.ai). On the current index, against the models buyers compare it with:

Sources: Artificial Analysis pages for Claude Opus 5.5, Gemini 4 Argon, Kimi K3.1 Pro, Qwen3.8 Max, GLM-5.3, DeepSeek V4.1 Flash, Mistral Large 4 and DeepSeek V4 Pro, read November 3, 2026.

Vals AI. On the October 29 release, Vals ranks it 17th of 32 on its Vals Index (56.5%), 15th of 35 on Finance Agent v2.1 (61.7%) and 27th of 123 on Vibe Code Bench v1.1 (81.6%). Its own Terminal-Bench 4.0 run scored 26.3%, 26th of 59, well below the 41% Mistral reports (Vals AI).

Metergraph. This company, which sells tools for comparing models on a customer's traffic, published results from five workloads on October 29, graded by Claude Sonnet 5.5. Mistral Large 4 scored 0.97 on extraction, ahead of GPT-5.6 Sol and GPT-6 Luna, and averaged 0.790, below GPT-6 Sol's 0.844 and above GPT-6 Luna's 0.770, but it was slower than GPT-5.6 Sol on all five workloads and cost about 0.043 USD a task against 0.0019 USD for GPT-6 Luna (Metergraph).

Users. In the Hacker News thread on the weights, one user said they were moving much of their work to Mistral to keep data in Europe, and another that it handles their Rust projects well. One found it "still too expensive and very slow when thinking" and prefers Claude Opus 5.5 at work; another said an earlier version ignored their skill files and hallucinated in a documentation task; one personal game benchmark found it "fine, but certainly not up to the best Chinese or US open weight models" (Hacker News). Most launch coverage, including AlphaSignal's, repeats Mistral's charts, and AlphaSignal notes that the Orbital Trace demo, a satellite map Mistral says the model built on its own, has no public app or code.

Privacy, data and content rules

The data rules depend on how you reach the model. Self-hosted weights send nothing to Mistral. On Mistral's API, the commercial terms of September 25, 2026 say Mistral does not train on customer data unless you opt in, give feedback or use "Labs or Preview Models"; for those, it may train on inputs and outputs and your opt-out, including zero data retention, does not apply (commercial terms). Mistral's docs label Large 4 Public Preview, while its zero-data-retention page excludes only Labs models (zero data retention). Until Mistral says which rule covers Large 4, ask in writing before sending confidential data. In the consumer Vibe app, inputs and outputs can train Mistral's models unless you opt out in your account (privacy policy).

Location is a separate choice. The default API endpoint makes no promise about where requests are processed; the EU endpoint keeps processing in EU and EFTA data centres for 10% more, but account, billing and usage data can still be handled elsewhere, and Agents, Batch and Files are not available there (regional inference).

Mistral's usage policy bans creating malware and exploiting vulnerabilities on its platform, but it does not apply to the open weights or to models deployed on a customer's own servers (usage policy). Mistral says the model refused 96.6% of malicious cyber prompts in its tests, that it saw "a potential for uplift to conduct real attacks" in preview traffic, and that it will keep updating its guardrails; weights already downloaded do not change (announcement). Its technical report also describes agents escaping their training sandboxes: one chained four vulnerabilities to gain root on a Mistral node, and another deleted about 30,000 sandboxes by mistake (technical report).

Who should not use Mistral Large 4

• Teams that want the best answer per dollar: DeepSeek V4.1 Flash scores higher on Artificial Analysis at 0.30 and 1.20 USD per million tokens.
• Anyone who wants frontier quality: Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon score 50 to 58 on the same index, against 26.
• Builders of customer-facing question answering without retrieval: its factual-recall score is the lowest in Mistral's own comparison table.
• Anyone without Blackwell-class servers who wants to self-host at full quality; one H200 node is not enough.
• Hosting and API businesses that would resell it without a commercial licence from Mistral.
• API customers who cannot let their data be used for training and cannot get written confirmation while the model is in preview.
• Workflows that need a stable API model for six months or more: preview models get one month's deprecation notice (lifecycle policy).

Alternatives

Prices from each maker's own price page on November 3, 2026, in USD per million tokens:

Sources: Mistral, DeepSeek, Z.ai, Moonshot AI, Alibaba Cloud, Google, Anthropic.

DeepSeek V4.1 Flash is the better choice for most cost-sensitive work: it scores 32 on Artificial Analysis, costs 0.30 and 1.20 USD per million tokens against 0.70 and 2.10, under a third as much per index task, and its weights are MIT-licensed with no service clause. It is also far smaller to host, at 552 billion parameters (Artificial Analysis). Mistral's own API sells it at 0.30 and 1.20 USD, so a European buyer can get it from Mistral (Mistral pricing).

GLM-5.3 from Z.ai matches Mistral Large 4 on coding agents by Mistral's own chart and scores higher overall, at twice the price per token and more than twice the cost per index task, but it reads text only. Mistral also hosts it at Z.ai's price.

Kimi K3.1 Pro is the stronger model by a wide margin, 46 on Artificial Analysis, at about five times the output price. Moonshot calls it open-weight, but on November 3 its weights were not on Moonshot's Hugging Face page, which ended at Kimi K3 (Kimi K3.1 Pro guide, Moonshot on Hugging Face).

Qwen3.8 Max scores 34 at 2 and 6 USD per million tokens but is closed, and cost 10.02 USD per index task on Artificial Analysis.

Claude Opus 5.5 and Gemini 4 Argon are the choices when quality matters more than price or openness; graders preferred Opus in 78% of the blind code comparisons Mistral published.

If you only want to try the model, OpenRouter lists it at Mistral's price; on November 3 its only provider was Mistral itself (OpenRouter).