
On October 6, 2026, Mistral AI opened a public preview of Mistral Large 4. The new AI model is now available for testing via the Mistral Studio API and, according to the company’s announcement, should be downloadable by the end of October. Developers can start experimenting today, while self-hosted deployments will have to wait for the model weights.
What’s new in Mistral Large 4?
Mistral Large 4 blends text and image processing with stronger reasoning. The model docs highlight document Q&A, function calling, and structured output. The latter lets apps receive answers in a predefined schema, not just free-form text.
The architecture lists 1.05 trillion parameters—the values the model learns during training—using a Mixture-of-Experts setup where only parts of the network activate per step. The official Hugging Face page cites 49 billion active parameters per token, or 52 billion including embedding and output layers. That’s why both figures appear across announcements.
The model weights aren’t out yet. Hugging Face points to October 31 as the expected release; Mistral says by month’s end. Open weights would enable running the model outside Mistral’s own stack. The final license will determine the usage terms. For background on the company and its models, see our guide to Mistral AI.
Preview pricing and context limits
The current Mistral price list shows a promotional rate of $0.68 per million input tokens and $2.09 per million output tokens. Cached input is listed at $0.07. Tokens are the chunks language models process; input and generated output are billed separately.
The crossed-out prices are $1.36, $4.18, and $0.14 respectively—so the lower numbers are clearly a discount. No end date is provided. If you’re budgeting, account for the higher rates as a conservative estimate.
Vercel also offers the model via its AI Gateway, without requiring a separate Mistral account. Existing apps can call the model through multiple supported API formats—an easy path to trial the preview in current workflows.
The context limit isn’t consistently stated. Mistral’s docs cite one million tokens, while Vercel’s model page lists 524,000 tokens for its connection. This window caps how much information a single request can include. For long case files or large codebases, the service’s limit matters more than the headline number on the model card.
Early benchmarks: promising, but read the fine print
Mistral spotlights coding, cybersecurity, and professional tasks. It reports a 93 percent score on Cybench, a set of forty security challenges from competitions. Until the weights land, Mistral says the model is being tested by security experts, selected partners, and government agencies.
A broader view comes from Artificial Analysis, which gives the preview 38 points on its Intelligence Index—a composite across ten evaluations spanning coding, scientific tasks, and document work. That score speaks to overall capability rather than a single niche benchmark.
Cost per completed task also matters. Artificial Analysis notes Large 4 generated a lot of output tokens in testing: 200 million versus a median of 81 million in its peer set. A low per-token price doesn’t automatically equal the lowest bill. For businesses, trials with real documents, target answer lengths, and verifiable outcomes are more telling than leaderboard ranks.
Next move: control your stack
Mistral says it trained Large 4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers. The public preview runs on the same infrastructure—tying model development directly to the capacity customers use.
That strategy has been in motion for a while. We previously covered Mistral’s Swedish data center plans. Partner EcoDataCenter confirmed a €1.2 billion investment with a planned 2027 opening. That future site is separate from the infrastructure powering today’s preview.
For organizations, the choice is clear: run via API and let the provider handle execution, or self-host and take control. With Large 4, technical evaluation can start now. A decision to deploy independently will hinge on the released weights, licensing terms, and a hard check on required hardware.
Popular news

AWS lets AI agents buy select tools on the developer’s cloud bill

Human clues helped AI agents mimic supply-chain decisions

Cline-Bench tests coding agents three times on real software tasks

ElevenLabs moves to buy Kotoba for its Japanese voice AI

Claude Code vs Cursor: Try the Same Fix in Both

