Mistral Large 4

Status: Public Preview. API access since October 6, 2026. Training still in progress at launch. Open weights and license not yet published as of October 24, 2026.

What is confirmed and what is not

Mistral Large 4 launched with a lot of promises attached, so it helps to separate what you can use today from what is still coming.

Available now: a public preview through Mistral Studio and the Mistral API, with chat completions, function calling, structured outputs, batch processing, document Q&A, built-in tools, and Mistral’s Agents and Conversations endpoints. Pricing is published.

Not yet available: the open weights, the license they will ship under, and the fuller architecture, benchmark, and post-training details Mistral said would follow. Mistral committed to releasing weights “by the end of the month,” meaning October 2026. Mistral’s Hugging Face placeholder points to October 31, while some reports cite October 27. Until the files and license appear, treat “open-weight” as a stated plan.

Mistral also said at launch that its reinforcement learning run was still in progress and that it expected further improvements. In other words, the model you test in the preview may not be the model that ships as final.

Specs

Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers.

Benchmarks, as reported by Mistral

Mistral published a long list of results, all of them its own unless noted. The coding numbers include 61.7% on DeepSWE v1.1 and 59.4% on SWE-Atlas-QnA. On AutomationBench, a set of 657 business workflows, it reports 59.9%. In cybersecurity, Mistral says Large 4 ranks in the top five on the Artificial Analysis Cyber Index and scores 82% on a vulnerability reproduce-and-patch test. It also notes that some closed models score near zero on that test because they refuse the task, which says as much about refusal policies as capability.

One result is worth singling out because someone else ran it. In a blind human coding evaluation run by Surge AI, on a 1 to 5 scale, Large 4 scored 3.74, second of five models tested, behind Claude Opus 5 at 4.22 and ahead of Kimi K3, GLM-5.3, and GLM-5.2. That is a small, specific test, but it is more informative than self-reported leaderboard positions.

Treat the rest as the company’s case for the model. Independent results on the final release will be the better guide, and the preview may change before then.

Running it and paying for it

In the preview, the only official route is Mistral’s API and Studio. Mistral says the model runs in multiple regions, including a Mistral-operated European deployment under European law, and that self-deployment, private cloud, and on-premise setups are supported for security customers.

API prices are $1.36 per million input tokens, $0.14 per million cached input tokens, and $4.18 per million output tokens. Mistral ran a 50% discount for two weeks after launch, which has likely ended. That places it above the cheap execution models like DeepSeek V4.1 Flash and MiniMax M3.1 Flash, and well below closed frontier APIs. As always, the useful comparison is cost per completed task, not the per-token rate.

Once weights are out, self-hosting a model this size will need substantial multi-GPU hardware. Only 52B parameters are active per token, but the full roughly 1T-parameter checkpoint still has to sit in memory.

Why it matters

Mistral is the leading European AI lab, and Large 4 is its first model at the trillion-parameter scale that the largest Chinese and US open-weight releases have reached. For organizations with EU data-residency requirements or a preference for a European supplier, it fills a real gap. It also adds competition in the large open-weight tier alongside Muse Spark Open, Kimi K3, GLM 5.2, and Qwen3.8-Max.

Whether it is the right choice depends on the same things as any model: how it does on your tasks, the license terms when they arrive, and the price you actually pay.

Beyond Large 4: Voxtral 3 and Voice Agents

Large 4 is Mistral’s flagship text and vision model, but it is not the whole company. On October 20, 2026, Mistral released Voxtral 3, a family of speech models aimed at voice agents. It is in public preview as well.

The family splits the job into parts. Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3 turn recorded audio into text, Voxtral Mini Transcribe Realtime 3 does live transcription, and Voxtral Speech 3 generates speech from text. Mistral says transcription covers 50+ languages; at least one review puts it closer to 35+ languages and 50+ dialects. Artificial Analysis ranked Voxtral Small Transcribe 3 third on its speech-to-text leaderboard, which is an independent result.

Preview pricing is $0.0035 per minute for Mini batch transcription, $0.004 for Small, $0.008 for realtime, and $13 per million characters for speech. Mistral also released a voice agent SDK, built on LiveKit, that chains speech recognition, a reasoning model, and speech output into one pipeline, with each stage swappable. Mistral reports turns completing in under a second; barge-in latency was not specified.

One change from earlier Voxtral releases: Voxtral 3 is closed-weight for now. It runs through Mistral’s API on European, US, and global endpoints, or through self-deployment arranged with Mistral, and Mistral has not said whether weights will follow. If open weights for speech matter to you, that is a real difference from the previous generation.

Taken together, Mistral’s 2026 releases point toward enterprise agents that combine text, vision, and voice, with a European deployment option. That is a reading of the product releases, not a strategy Mistral has spelled out in full.

Limitations

• Preview, not final. Behavior and quality may change before general release.
• Weights and license pending. You cannot self-host yet, and the license terms could affect commercial use.
• Mostly vendor benchmarks. Only a few results so far come from outside parties.
• Context limits vary by endpoint. Check the actual limit on your provider.
• Pricier than budget models. It sits in the middle of the market, not at the low end.
• Voxtral 3 is closed-weight and in preview. Mistral’s new speech models are not downloadable, unlike earlier Voxtral versions.

Alternatives

• Muse Spark Open, Meta’s 816B open-weight model with published weights
• DeepSeek V4-Pro, a large MoE model at lower API prices
• Kimi K3 and GLM 5.2, other large open-weight models with 1M context
• OpenRouter, to test several of these through one API
• ElevenLabs and Gemini 3.8 Live, alternatives to Voxtral 3 for voice agents

Background: Open Models vs Closed Models, open weights, and AI vendor lock-in.

Compare: Muse Spark Open vs Mistral Large 4.