The trillion is a routing number
Large 4 routes each token through 49 billion active parameters out of 1.05 trillion total. That is the standard mixture-of-experts arrangement: the headline counts every expert in the building, the compute bill counts the ones that actually fire for a given token. Mistral's docs list the API identifier mistral-large-4, structured outputs, function calling, document QA, and agent tooling — the plumbing for connecting a model to software workflows rather than chatting with it.
Preview pricing ($0.68 in, $2.09 out per million tokens, roughly doubling when the preview ends) charges for the active slice, not the trillion. The GPU-memory estimate for self-hosting is currently unspecified in the documentation, and Mistral says the model is still training, so preview results are a moving snapshot. Teams evaluating it for production should version-pin and retest rather than trust day-one numbers.
Vendor benchmarks, read with the harness in mind
Mistral's published scores: 61.7% on DeepSWE 1.1, 28.3% on Terminal-Bench 4, 59.9% on AutomationBench, 82% on reproducing and patching a real vulnerability, 93% on Cybench's security challenges. The comparisons it chose are DeepSeek, Qwen, and Kimi, and the results vary by workload. Artificial Analysis puts the preview at 38 on its Intelligence Index.
Coding outcomes move with the agent harness, the tool budget, and the evaluation rules. A vulnerability-reproduction score moves with something simpler: whether the model was permitted to attempt the task. These are Mistral-run snapshots of a model that is still being trained. They are a starting point for independent testing, not a verdict.
The superlative with its rivals defined out of it
The launch copy calls Large 4 “the best open weights model from US or Europe on aggregated benchmarks.” Read that twice. The claim excludes the Chinese labs, then adds that Large 4 is “competitive with” the strongest open-weight models from China, on Mistral's own aggregated comparisons. It is a superlative that defines its main rivals out of the ranking, then borrows their credibility back in the next sentence.
“Open” is currently a calendar entry
The model card marks both weights and license “coming soon,” with the stated target the end of October. Developers can test the hosted preview today; they cannot download anything, and the terms they would download it under are unpublished. Mistral also says the model was trained “from scratch,” a pointed choice of words in a season when Anthropic and OpenAI have accused Chinese labs of distilling their models.
Contrast this with yesterday's Kandinsky 6.0: code, checkpoints, and diffusers integration shipped under plain MIT, downloadable now. Mistral shipped a date. An open-weight model with unpublished weights and an unpublished license is a press release with a schedule.
The part to read twice
Vetted cybersecurity partners and state authorities are already testing the same model with reduced moderation and expanded cyber capabilities, before the public weights release. Mistral's argument is that provider-level refusals get in the way of legitimate vulnerability research and incident response. That is plausibly true. It also means the least-restrained version of Large 4 is already circulating among selected organizations while the downloadable one does not exist yet. If you are keeping score on how frontier capabilities diffuse, this is the detail that matters more than the parameter count.
Sovereignty is infrastructure, not a benchmark
Mistral trained Large 4 on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers and serves the preview from that infrastructure. For a European buyer, where data is processed and who controls service access are real questions, and Mistral is selling an answer to both. None of that makes the model more accurate. Judge the sovereignty pitch on procurement terms and the model on independent evals.
Sources
- [1] AIstify — “Mistral Unveils Large 4 ‘Le Chonk,’ a Trillion-Parameter AI Model” (Oct 6, 2026)Read source
- [2] Unite.AI — Artificial Intelligence news and analysisRead source