Mistral Large 4 lands as a 1T-parameter API preview, and independent evals put it at mid-tier intelligence for a mid-tier price

Mistral opened a public preview of Mistral Large 4, a 1-trillion-parameter mixture-of-experts model with 49 billion parameters active per token, and says the weights will follow by the end of October. The vendor’s own benchmark table is ambitious, but the more useful signal arrived the same day from outside the company: Artificial Analysis tested the preview API and placed it level with GPT-6 Luna on its Intelligence Index. That makes Large 4 a credible new option, though not a frontier-class one, and its price per completed task matters as much as its score.

What Mistral shipped

Mistral describes Large 4 as a natively multimodal hybrid model that handles both instruct and reasoning modes, trained on Nvidia Grace Blackwell GPUs in European datacenters. The preview API is live in Mistral Studio at $1.36 per million input tokens and $4.18 per million output tokens.

Two details stand out for teams evaluating it:

• Deployment. Mistral says the model will run in multiple regions, including a European deployment it operates end to end, independently of other digital service providers and under European law.
• Weights later, terms unknown. The announcement commits to releasing weights by the end of the month but does not name a license. Mistral Large 3 shipped under Apache 2.0, so the license for Large 4 is the open question that decides whether self-hosting is a practical path.

For developers, the immediate product is an API. The open-weight story is a promise with a date attached, not something you can download today.

What the independent numbers say

Artificial Analysis scored the preview at 38 on its Intelligence Index, which aggregates ten evaluations. That puts it alongside GPT-6 Luna (38) and DeepSeek V4.1 Flash (39), and Artificial Analysis calls it the most intelligent model from outside the US and China. The Decoder, citing the same index, notes that Claude Opus 5.5 scores 58. Large 4 is a real step for Mistral, but the gap to the top closed models remains wide.

Security work is where the preview looks strongest. Artificial Analysis gave it 50 on its Cyber Index, level with GLM-5.3-Flash, and 82% on CyberGym-E2E-AA, a test where a model reproduces a real vulnerability in open-source software and then patches it. That edges out MiMo-V2.6-Pro (79%) and GPT-6 Luna (78%). Mistral, for its part, says the model refuses cyber prompts more often than other open models, a claim that independent testers have not yet separated into legitimate and malicious requests.

One limit on the evidence: every independent number so far comes from the hosted preview, and outside benchmarks against downloadable weights will not exist until the weights ship.

The cost per task problem

Token prices only tell part of the story. Artificial Analysis puts the cost of running one Intelligence Index task on Large 4 at $1.13, or $0.57 during a two-week 50% launch discount. DeepSeek V4.1 Flash, at roughly the same score, costs $0.27 per task, and GLM-5.3-Flash costs $0.25. At list price, Large 4 delivers comparable measured intelligence for about four times the spend of those Chinese open-weight peers.

That gap reframes who should care. Teams choosing purely on capability per dollar have cheaper options at this tier. Large 4’s case rests on where it runs and who operates it rather than raw efficiency.

What it means for developers

The practical audience is narrower than the “1T parameters” headline suggests, but real:

• EU-regulated teams that need a model operated in Europe, under European law, by a European company now have a noticeably stronger option than Mistral offered a month ago.
• Security and AppSec teams testing models for vulnerability reproduction and patching should add Large 4 to their evaluation set, with the caveat that refusal behavior on borderline requests is still undocumented.
• Teams planning to self-host should wait for the license and the weights before committing. A 1T MoE with 49B active parameters is also a serious hardware commitment even when the license is permissive.

Cost-sensitive workloads that do not need EU operation are better served by benchmarking against DeepSeek and GLM first. For them, Large 4 is a model to watch rather than a default.

What to watch

The next signals arrive within weeks: the license attached to the weights, whether the end-of-October date holds, and independent results on the downloadable model rather than the hosted preview. If the weights ship under Apache 2.0 at this capability level, the cost comparison changes for anyone with their own GPUs. Stackmaven’s follow-up will land after the weight release.

Sources cited

1. Introducing Mistral Large 4 (Mistral AI) mistral.ai
2. Mistral has released Mistral Large 4, making France home to the most intelligent model outside the US and China (Artificial Analysis) artificialanalysis.ai
3. Mistral Large 4 is Europe's trillion-parameter answer to US models that refuse security work (The Decoder) the-decoder.com
4. Mistral's new 1T model aims to leapfrog closed and open rivals (TechCrunch) techcrunch.com