Mistral Large 4
Status: Released with open weights. Public API preview from October 6, 2026; weights published on Hugging Face on October 29, 2026 under the Mistral Ecosystem Open-Weight License.
What changed with the October 29 release
Mistral Large 4 spent most of October as an API-only preview with a promise attached: weights “by the end of the month.” Mistral kept that promise on October 29. The checkpoint, mistralai/Mistral-Large-4-1T-A52B, is a free download on Hugging Face with no approval step, alongside FP8 and an NVFP4 build aimed at NVIDIA Blackwell GPUs.
The release model is not quite the preview model. Mistral finished the reinforcement learning run that was still going at launch, and the numbers it published moved accordingly: Terminal-Bench 4.0 went from 28% in the preview to 41%, with about 30% fewer output tokens. The API price also came down, to $0.70 per million input tokens and $2.10 per million output, roughly half the preview list price.
You can now get the model in several ways: Mistral Studio and the Mistral API (mistral-large-4), OpenRouter, Microsoft Foundry with private Azure deployment options, NVIDIA-hosted endpoints and a NIM container, and the raw weights for your own servers.
Specs
| Mistral Large 4 | |
|---|---|
| Developer | Mistral AI |
| Released | Preview October 6, 2026; open weights October 29, 2026 |
| Architecture | Mixture-of-Experts, 128 routed experts plus one shared expert, 4 routed experts active per token |
| Parameters | About 1T total (1.05T in Mistral’s docs); 49B active per token (about 52B counting embedding and output layers) |
| Vision encoder | 1.6B parameters, Pixtral-style |
| Context window | 1M tokens |
| Input | Text and images, including charts and documents |
| Reasoning | Three settings: low, medium, and high (the default) |
| Languages | 160+ per Mistral |
| Weights | Hugging Face, including FP8 and NVFP4 variants |
| License | Mistral Ecosystem Open-Weight License (modified MIT, with a Model-as-a-Service restriction) |
Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers.
A trillion parameters, but only 49 billion at a time
The headline number needs some unpacking. In a Mixture-of-Experts model, most of the network is split into many specialist sub-networks (“experts”). For each token, a router picks a handful of them. Large 4 has about 1 trillion parameters in total, but only around 49 billion do work on any given token.
That makes generation much cheaper than a dense trillion-parameter model would be: compute per token is closer to that of a 50B model. It does not shrink the model you have to load. The router can pick any expert for the next token, so every expert has to stay in fast GPU memory. At 8-bit precision, a trillion parameters is roughly a terabyte of weights, before you add the memory needed for long contexts.
You will see slightly different figures quoted. The Hugging Face model card says about 1 trillion total and 49 billion active (52 billion once the embedding and output layers are counted), while Mistral’s model documentation gives 1.05 trillion total and 52 billion active, and lists a separate 1.6-billion-parameter vision encoder. These describe the same model counted in different ways, not different models.
So there are three different numbers to keep apart. Total parameters tell you how much memory the model needs. Active parameters tell you roughly how much compute each token costs. The hardware you need is driven by the first number, not the second. “Only 49B active” does not mean it fits on a GPU that could run a 49B dense model.
Self-hosting in practice
Mistral’s model card shows every serving example (vLLM, SGLang, Transformers) with eight-way tensor parallelism, which means one model split across eight data-center GPUs in a single server. The NVFP4 variant cuts the memory footprint on Blackwell hardware. Community GGUF builds for llama.cpp, and through them Ollama and LM Studio, offer smaller quantized versions, but even aggressive quantization leaves a model far beyond a consumer graphics card, and quality degrades as you compress. For fine-tuning, the model card lists Axolotl, Unsloth, and TRL.
The weights are also only the start. A production deployment needs a serving stack, monitoring, scaling, security patching, and someone who understands GPU inference well enough to tune batch sizes and context limits. For most teams, the API at $0.70 / $2.10 will cost less than owning eight-GPU servers until usage is large and steady. See when local AI makes sense for the trade-off in general terms.
Where self-hosting does earn its cost is control. A bank, a government department, a healthcare provider, or a company with strict confidentiality rules can run Large 4 inside its own network, so prompts and documents never leave infrastructure it controls. It can pin a version, fine-tune on internal data, and stop depending on a vendor’s API staying available (AI vendor lock-in is a real cost). Self-hosting does not make a deployment compliant or secure by itself, though. Access controls, logging, data handling, and model risk reviews are still the organization’s job.
The license: downloadable is not the same as unrestricted
Large 4 does not use Apache 2.0 like Mistral Large 3 did. It ships under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT license. In practical terms:
- Downloadable: anyone can get the weights from Hugging Face without an approval step.
- Self-hostable: you can run, modify, and fine-tune the model on your own infrastructure.
- Redistributable: you can pass on the original or modified weights, as long as the copyright and permission notice go with them. Recipients get the same terms, including the clause below.
- Commercially usable: commercial use is broadly allowed, for example using the model inside your own product or internal tools. Internal use that does not make the model available to third parties is explicitly exempt from the extra license.
- Model-as-a-Service is carved out: if you or an affiliate run a business that sells hosted access to the model, you need a separate commercial license from Mistral first. The license defines this as giving third parties inference or post-training access, for example through an API or hosted endpoint, where they control the inputs, parameters, or training data. End-user products that use the model inside specific features are excluded from that definition, and so is simply passing requests to a model someone else hosts.
That carve-out is why commentators point out that the license is not OSI-approved open source, and Mistral did not release its training data or full training pipeline either. The license sets no revenue threshold and contains no cybersecurity-specific clauses. The distinction that matters for most organizations is between using the model to run your own business and reselling the model as an API. Cases in between, such as a product whose main feature is thin access to the model, are where you should read the license text and talk to a lawyer rather than rely on a summary like this one. See open weights and open source AI for the general distinction.
The restricted cybersecurity variant
Alongside the public release, Mistral is offering preview access to a separate variant tuned for offensive security work: penetration testing, vulnerability assessment, and red-teaming. It is not in the Hugging Face download. Access goes to enterprise design partners such as Orange Cyberdefense, after vetting and under contractual safeguards, and use is limited to systems the partner owns or is explicitly authorized to test.
The general model has its own cyber safeguards. Mistral reports a 96.6% average refusal rate on malicious cyber requests, which it says is the highest it measures among open models, and 97.1% attack resistance on the B3 Agent Security Benchmark, up from 93.3% in the preview. Mistral has also said its safeguards still need adjustment, so treat these as the company’s own measurements.
Benchmarks, as reported by Mistral
Mistral calls Large 4 the best open-weight model for coding. Its main supporting figures, all published by Mistral:
- Artificial Analysis Coding Agent Index: 48%, which Mistral describes as effectively tied with GLM-5.3, while using 47% fewer tokens per task on average. Fewer tokens for the same result matters directly for cost per completed task.
- Terminal-Bench 4.0: 41%, up from 28% in the preview.
- FrontierCode 2.0 (Cognition’s private benchmark): 19.7%.
- BrowseComp: 84.5%. FinWorkBench: 73.3%, which Mistral says is second among open models only to Kimi K3.
- Human preference (Surge AI coding evaluation): preferred over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, behind only Claude Opus 5.5 among the models tested.
The Artificial Analysis and Surge AI results come from outside organizations, which makes them more useful than self-run leaderboards, but Mistral chose which results to publish. A tie on one coding index does not make Large 4 the best model for your workload. Test it on your own tasks before switching.
Pricing
| Per 1M tokens | |
|---|---|
| Input | $0.70 |
| Cached input | $0.07 |
| Output | $2.10 |
These are Mistral API prices for the released model. That puts it well below closed frontier APIs and among the mid-priced open models, though still above the cheapest execution models such as DeepSeek V4.1 Flash and MiniMax M3.1 Flash. Prices on OpenRouter, Microsoft Foundry, or NVIDIA endpoints can differ. Self-hosting has no per-token fee, but you pay for hardware, power, and operations instead.
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 that want a capable open-weight model from outside the US and China, it is now one of the few downloadable options at this size. It also adds competition in the large open-weight tier alongside Muse Spark Open, Kimi K3, GLM 5.2, and Qwen3.8-Max. For the wider picture, see frontier open models explained.
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.
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. That contrasts with Large 4, whose weights are now public.
Limitations
- Heavy to run yourself. Downloadable does not mean easy: plan for a multi-GPU server, not a workstation.
- License restrictions on hosting. Reselling the model as a service needs a separate agreement, and the license is not OSI open source.
- Mostly vendor-selected benchmarks. A few results come from outside organizations, but independent testing of the final release is still early.
- Context limits vary by endpoint. The model supports 1M tokens, but providers can set lower limits.
- The cyber variant is not public. Only vetted partners can access it.
- 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
- DeepSeek V4-Pro, a large MoE model at lower API prices
- Kimi K3 and GLM 5.3, other large open-weight models with 1M context; see Kimi K3.1 Pro vs Mistral Large 4 and Mistral Large 4 vs GLM-5.3
- OpenRouter, to test several of these through one API
- ElevenLabs and Gemini 3.8 Live, alternatives to Voxtral 3 for voice agents, and NVIDIA Nemotron 3.5 Transcribe if you need open-weight transcription of recorded audio
Background: Open Models vs Closed Models, Local AI explained, and AI vendor lock-in.
Compare: Muse Spark Open vs Mistral Large 4, Mistral Large 4 vs DeepSeek V4.1 Flash, and GPT-6.1 Sol vs Mistral Large 4 vs Kimi K3.1 Pro. For how Mistral fits into Europe’s AI industry, see USA vs Europe vs China: the AI models and tools compared.
Continue learning
Explore related guides, tools, workflows, and prompts that help you go deeper into this topic.
See how this tool fits into a workflow
Browse step-by-step AI workflows that use ChatGPT, Claude, Gemini, and other tools.
Frequently Asked Questions
Can I download Mistral Large 4?
Yes. Mistral published the weights on Hugging Face on October 29, 2026, after a public API preview that began October 6. Besides the main checkpoint there are FP8 and NVFP4 variants, and the model card lists vLLM, SGLang, Transformers, and llama.cpp as supported runtimes.
Is Mistral Large 4 open source?
It is open-weight, but not open source in the strict sense. The weights ship under the Mistral Ecosystem Open-Weight License, which Mistral calls a modified MIT license. Use, modification, fine-tuning, and commercial use are broadly allowed, but a business that offers the model as a hosted service (Model-as-a-Service) needs a separate commercial license first. Training data and the full training pipeline were not released. Mistral Large 3 used Apache 2.0, so do not assume the old terms carry over.
Can my company use Mistral Large 4 commercially?
For most companies, yes. The license exempts internal use that does not make the model available to third parties, and its Model-as-a-Service definition excludes end-user products where the model powers specific features. What needs a separate agreement with Mistral is selling access to the model itself: an API or hosted endpoint where your customers control the prompts, parameters, or training data. Products that sit between those two are a legal question, so read the license text and get legal advice for anything borderline.
Can I run Mistral Large 4 on my own GPU?
Not on a normal consumer GPU. Only about 49 billion parameters are active for each token, but all of the roughly 1 trillion parameters have to be loaded into memory. Mistral's own serving examples split the model across eight GPUs. Heavily quantized community builds exist, but they still need far more memory than a gaming PC or laptop has, and quality drops as you compress further.
How big is Mistral Large 4?
About 1 trillion total parameters in a Mixture-of-Experts design with 128 routed experts and one shared expert. The model card gives 49 billion active parameters per token, or about 52 billion when the embedding and output layers are counted, which is why both numbers circulate. The context window is 1M tokens.
What is the cybersecurity version of Mistral Large 4?
Separately from the public weights, Mistral is giving preview access to a specialized variant tuned for penetration testing, vulnerability assessment, and red-teaming. It is not part of the open download. Access goes to enterprise design partners such as Orange Cyberdefense, after vetting and under contract, and testing is limited to systems the organization owns or is explicitly authorized to test.
What is Voxtral 3?
Voxtral 3 is Mistral's speech model family, released in public preview on October 20, 2026. It includes batch and realtime transcription models and a speech generation model, plus a voice agent SDK. Unlike earlier Voxtral releases, the weights are not currently published.
Last updated: