What it is
Mistral Large 4 is Mistral AI's flagship, announced on October 6, 2026 as a public preview on Mistral Studio under the API ID mistral-large-4. It is a mixture-of-experts model with about 1.05 trillion parameters, 52 billion active per token and a 1.6-billion-parameter vision encoder, and it combines quick replies and step-by-step reasoning in one model. It reads text and images and covers more than 160 languages, including every official EU language.
The preview costs $1.36 per million input tokens, $0.14 for cached input and $4.18 for output. Mistral's docs list a 1M-token context window; Artificial Analysis lists 524K. Mistral says the weights will be released by the end of October 2026 and the docs tag the model 'Open', but no licence has been named and, as of October 22, no Large 4 weights were on Mistral's Hugging Face page. Mistral also says the reinforcement-learning run is still in progress, so the model will change during the preview.
It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centres, and Mistral will offer a Europe-only deployment run end to end by Mistral under European law.
Our review
Large 4 is a sovereignty purchase first and a capability purchase second. Artificial Analysis gives the preview an Intelligence Index of 38, 65th of 214 models it tracks: comfortably above the median of 27 for comparable models, but well short of the US and Chinese leaders. Mistral's launch post leans on narrower wins instead, and most of those figures are its own.
The self-reported numbers show where it is strong. Mistral reports 82% on a reproduce-and-patch vulnerability test within Artificial Analysis's cyber index, where Claude Opus 5.5 and GPT-6 Astra scored near zero because they refused, and 59.9% on AutomationBench's 657 business workflows. Agentic coding is weaker: 28.3% on Terminal-Bench 4 and 3.74 out of 5 in Surge AI's blind coding evaluation, against 4.22 for Claude Opus 5.
The headline price hides a verbosity bill. Artificial Analysis measured 200 million output tokens to run its index, against a median of 87 million, ranking it 108th of 115 for wordiness. At $4.18 per million output tokens that cancels part of the saving over pricier models on reasoning-heavy work, and output speed of about 80 tokens a second is middling.
The open-weight promise is the part to watch. At roughly 1.05 trillion parameters, even 8-bit weights would need about a terabyte of GPU memory, so self-hosting is a rack-scale project for a ministry or a bank, not a laptop experiment. And the context window is listed as 1M tokens by Mistral and 524K by Artificial Analysis, a discrepancy worth settling with Mistral before designing around it.
Where it earns its place
A procurement lead at an EU public agency with a data-residency mandate
Buy a frontier-class assistant whose provider, hosting and law are all European
Use Mistral's EU-only deployment, which Mistral runs without third-party cloud providers. Contract terms and SLAs come through Enterprise sales; the preview label means production commitments should wait for general availability.
A security team at a European bank running internal red-team exercises
Have a model reproduce and patch known vulnerabilities rather than refuse
Mistral reports 82% on a reproduce-and-patch test where US frontier models mostly refused. Requests run through the API at $1.36/$4.18 per 1M tokens; vetted partners get a lighter-moderation build before the weights ship.
A platform engineer at a multilingual e-commerce marketplace
Classify and answer support tickets in 20-plus European languages
One million replies a month at about 1,500 input and 400 output tokens each costs roughly $3,700 at list price, or about $1,850 through the batch API. Turning reasoning down matters, because the model's default verbosity inflates the output bill.
Before you commit
- Setup
- Create a Mistral Studio key and call mistral-large-4 on /v1/chat/completions, /v1/conversations or /v1/agents; it is also listed on OpenRouter at the same price.
- Integrations
- Function calling, structured outputs, document Q&A, built-in tools, agents and a batch endpoint on Mistral's API; OpenRouter for multi-provider routing. AWS and Azure availability not announced.
- Data and privacy
- On Studio's free mode Mistral may train on inputs and outputs unless you opt out; pay-as-you-go customers can opt out. Zero data retention is granted on request at Mistral's discretion and does not cover batch, agents or conversations. Certifications are not stated on the pages we read.
- Cost at scale
- A million support replies at 1,500 in and 400 out tokens is about $3,700 a month at list, $1,850 batched; cached prompts cut input to $0.14 per 1M. Self-hosting the promised weights needs roughly 1TB of GPU memory at 8-bit, our arithmetic from the 1.05T parameter count.
- Limits
- Public preview with training still in progress; context 1M per Mistral docs, 524K per Artificial Analysis; output cap and rate limits not published; about 80 tokens a second and 2.3 seconds to first token on Mistral's API.
What's good
- Trained and hosted in Europe, with a Mistral-run EU-only deployment
- $1.36/$4.18 per 1M tokens, with 90% off cached input and 50% off batch
- Over 160 languages, including every official EU language
- Strong self-reported security and business-automation scores
What's not
- Intelligence Index of 38, 65th of 214 on Artificial Analysis
- Very verbose: 200M output tokens on the index against an 87M median
- Still a preview; weights and licence not released as of October 22
- Context window reported as 1M by Mistral and 524K by Artificial Analysis
Best for
European organisations that need a frontier-class model with data sovereignty and a path to self-hosting.
Alternatives
Mistral Large 4: common questions
- Is Mistral Large 4 worth it?
- Europe finally has a trillion-parameter model of its own; independent tests place it in the upper middle of the pack, not at the frontier. We rate it 3.9 out of 5.
- How much does Mistral Large 4 cost?
- Preview API: $1.36 per 1M input tokens ($0.14 cached), $4.18 per 1M output; batch 50% off. Open weights promised by end of October 2026.
- What are the downsides of Mistral Large 4?
- Intelligence Index of 38, 65th of 214 on Artificial Analysis; very verbose: 200M output tokens on the index against an 87M median; still a preview; weights and licence not released as of October 22; context window reported as 1M by Mistral and 524K by Artificial Analysis.
- Who is Mistral Large 4 best for?
- European organisations that need a frontier-class model with data sovereignty and a path to self-hosting.
- How much does Mistral Large 4 cost for a team or in production?
- A million support replies at 1,500 in and 400 out tokens is about $3,700 a month at list, $1,850 batched; cached prompts cut input to $0.14 per 1M. Self-hosting the promised weights needs roughly 1TB of GPU memory at 8-bit, our arithmetic from the 1.05T parameter count.
- Is Mistral Large 4 safe for company data?
- On Studio's free mode Mistral may train on inputs and outputs unless you opt out; pay-as-you-go customers can opt out. Zero data retention is granted on request at Mistral's discretion and does not cover batch, agents or conversations. Certifications are not stated on the pages we read.
- What are the best Mistral Large 4 alternatives?
- The closest alternatives we have reviewed are Mistral Medium 3.5, DeepSeek-V4-Pro, GLM-5.3 and Claude Opus 5.5.
Answers checked Oct 22, 2026.