Mistral Large 4 Is Generally Available — a 1T-Parameter Open-Weight Model You Can Download
Europe's biggest open-weight model just crossed the line from preview to production-ready. Mistral Large 4 — nicknamed "Le Chonk" — is now generally available, and unlike most flagship models, you can download the whole thing.
It entered public preview on October 6, 2026; Mistral released an improved set of open weights on October 29; and its docs card has marked it generally available since November 5. The headline numbers: 1.05 trillion total parameters, 52 billion active per token, a 1M-token context window, and — according to Mistral — a 13-point jump on the Terminal-Bench coding benchmark after three extra weeks of reinforcement learning, while using 30% fewer output tokens.
What actually shipped, and when
The timeline matters, because the model changed between its first appearance and GA.
- October 6, 2026 — public preview: Large 4 launched as an API preview. Mistral's own Terminal-Bench 4.0 score at that point was 28.3%.
-
October 29, 2026 — open weights: Mistral released the full weights on Hugging Face (
mistralai/Mistral-Large-4-1T-A52B) in FP8 plus an NVFP4 checkpoint, with vLLM serving support. This checkpoint is not the preview model — it had three more weeks of reinforcement learning baked in. - November 5, 2026 — generally available: Mistral's model documentation switched the status from Public Preview to GA, and the weights release became the GA model.
- October 31, 2026 — the old flagships retired: Mistral deprecated both Large 3 and Medium 3.5, pointing users at Large 4. Both endpoints retire on April 30, 2027, so anyone still on them has a migration deadline.
The model itself is a mixture-of-experts design: 1.05T total parameters, but only 52B active on any given token, which is why it can be served at a mid-tier price despite the headline size. It takes text and images in, returns text, and supports structured outputs, function calling, document Q&A, batching, and Mistral's Agents API. Training data spans more than 160 languages, including every official EU language, and Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers.
The benchmark numbers — and the honest asterisk
Mistral's headline claim is that Large 4 is now the best open-weight model for coding. The numbers it cites:
- Terminal-Bench 4.0: 28.3% → 41% — a 13-point gain over the preview, with 30% fewer output tokens on that benchmark. Mistral credits the three extra weeks of reinforcement learning.
- Artificial Analysis Coding Agent Index: 48% — effectively tied with GLM-5.3, while using 47% fewer tokens per task on average, according to Mistral.
- BrowseComp: 84.5%, per Mistral's release notes.
Here is the asterisk, and it is not small. Independent evaluations of the same released weights score it well below Mistral's self-reported 41%: Artificial Analysis's own Terminal-Bench 4.0 page lists Large 4 at 24.7% (high effort) and 23.2% (medium), and Vals.ai's harness scores it below the preview too. The gap is a harness and setup difference — different agent scaffolding, different attempt counts, different reasoning effort — not proof that anyone is wrong. But it does mean you should not plan a migration on 41%. The one claim independent testing backs up in direction, if not in size, is the efficiency gain: output tokens per task really did fall sharply, and Artificial Analysis measured the cost per Intelligence Index task dropping from $5.54 to $2.06. Cheaper per task is the more reliable win here than the headline score.
What it costs on the API
Mistral cut the price when the weights shipped: $0.70 per million input tokens, $0.07 per million cached input, and $2.10 per million output — down from $1.36 and $4.18 at preview. That undercuts the model it replaces (Large 3 was $0.50/$1.50, but Medium 3.5 was $1.50/$7.50) and sits well below frontier closed models for agentic workloads where output tokens dominate the bill. The 30% output-token reduction compounds that: fewer tokens per task at a lower price per token. The API model ID is mistral-large-4. It is also available through OpenRouter, Microsoft Foundry, and NVIDIA-hosted endpoints.
What it means if you build with it
1. The API is the sensible default; the weights are for a specific kind of team. Downloading a 1T-parameter model is not a weekend project. Self-hosting recipes call for roughly an 8× Blackwell-class node, and the weights ship in FP8/NVFP4 for that reason. The weights matter if you need to fine-tune on your own data, run inside your own compliance boundary, or serve offline — the classic European enterprise and government use cases Mistral is built for. For everyone else, the API gives you the same model without the hardware bill.
2. Agentic coding is the real test, not chat. The gains Mistral reports are concentrated in terminal-based agent work: Terminal-Bench measures whether a model can finish real tasks in a command-line environment. If your use case is coding agents, SWE-style automation, or long-running terminal sessions, Large 4 deserves a benchmark run against whatever you use today — with your own harness, given the scoring gap above. If your use case is chat or summarisation, the cheaper models in Mistral's own lineup will do.
3. Read the licence before you build a business on it. The weights are under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT licence: free to use, modify, fine-tune and commercialise. But model-as-a-service businesses need a separate commercial licence. If your plan is to host Large 4 and sell API access to it, that clause applies to you — check it before you invest in the hardware.
4. If you are on Large 3 or Medium 3.5, the clock is running. Both were deprecated on October 31, 2026 and retire April 30, 2027. Large 4 is the named replacement for both, and it is cheaper than Medium 3.5 was. Migration is not urgent today, but it is a dated deadline — put it on the roadmap now rather than discovering it in April.
5. Europe's infrastructure bet is the backdrop, not the news. Mistral has committed roughly €4 billion to data centers in France and Sweden — a plan announced in May 2026, targeting 200 MW of capacity by 2027 and 1 GW by 2030 — and raised a €3 billion Series D in September at a valuation above €21 billion. That is why the model exists as an open-weight European flagship rather than a closed API-only product. It is context for the strategy, not a new announcement, and it does not change what the model can do for you today.
What we don't know yet
- How it holds up in independent hands: the gap between Mistral's 41% and independent harnesses' ~25% on Terminal-Bench needs a careful third-party replication before anyone quotes a single number with confidence.
- Real-world agent reliability: benchmark scores are not production uptime. The model's behaviour across long, multi-step agent runs — tool-call reliability, error recovery, cost per completed task — is still mostly unmeasured outside Mistral.
- Fine-tuning quality: the weights are out, but the community's LoRA and full fine-tune results are only just appearing. Whether the open model fine-tunes as well as it serves is an open question.
- Whether the price holds: the October price cut was steep. API pricing for open-weight flagships has moved before, and a model this large is expensive to serve — watch the rate card.
Sources: Mistral AI, Mistral Large 4 public preview announcement (October 6, 2026) and open-weights release (October 29, 2026); Mistral model documentation (GA dated November 5, 2026); Mistral API pricing page ($0.70 / $0.07 / $2.10 per million tokens); Artificial Analysis Terminal-Bench 4.0 and Coding Agent Index pages; Vals.ai Terminal-Bench 4.0 leaderboard; VentureBeat / CNBC reporting on Mistral's €4 billion data-center commitment (May 2026) and €3 billion Series D (September 8, 2026). All benchmark figures are as reported by Mistral or the named independent evaluators; the 41% Terminal-Bench figure is Mistral's self-reported number and differs from independent harnesses. No hands-on testing was performed for this article.
This article was first published on ToolsDig.
Top comments (0)