On Tuesday, the rumor became a product — this time from Paris. Mistral AI unveiled Large 4, its new flagship open-weight model: 1.05 trillion total parameters, a granular mixture-of-experts design, and a 1-million-token context window. Industry insiders have already christened it “Le Chonk.” It is Europe’s boldest answer yet to the open-model wave led by China’s DeepSeek and Qwen — and for readers of this site, it raises the question we get asked most: does “open weights” mean “uncensored“?
Short answer: not exactly. Longer answer below — plus what we’ll test the moment the weights actually drop.
TL;DR — What We Know
| Confirmed | Not yet |
|---|---|
| Model name: Mistral Large 4, announced Oct 6, 2026 (public preview via API) | Weights not yet released (planned Oct 27) |
| 1.05T total params, ~49B active per token (MoE), 1.6B vision encoder, 1M context | Custom license — departs from Large 3’s Apache 2.0; terms unknown |
| Preview API live: $1.36 / $4.18 per 1M input/output tokens | Pre-release: cybersecurity experts and governments get a less-restricted version before Oct 27 (Reuters) |
| Multimodal (text+image in, text out), 160+ languages | Benchmarks are company-reported, not independently verified |
| Trained ~2 months on ~3,800–4,000 Nvidia Grace Blackwell GPUs in Mistral’s EU data centers | Pricing/availability of private-cloud and on-prem options not fully detailed |
| Artificial Analysis: 38 on Intelligence Index (vs 9 for Large 3; ~20 behind Claude Opus 5.5) | No hands-on refusal testing possible yet — that’s our job when weights land |
What Mistral Claims
Mistral’s pitch isn’t just “we beat everyone.” It’s sovereignty. After the US restricted international distribution of select proprietary models in June, Mistral is selling Europe’s alternative: open weights you can run and customize locally, paying only for the compute you consume. Co-founder Guillaume Lample put it bluntly — with a closed model, “there is no guarantee it will still be there tomorrow.”
On the company’s own scoreboard: the strongest open-weight model from the US or Europe “by a substantial margin,” leading open models on SciCode-Verified (scientific coding) and ahead of DeepSeek V4 Pro and Qwen3.8 Max on a combined coding-agent index. The efficiency story is real too — VP of Science Pierre Stock says training burned roughly two to three times less compute than Chinese competitors. Treat every one of these numbers as provisional: the weights aren’t public yet, so nobody outside Mistral can reproduce them, and independent analysts note the margins against Qwen3.8 Max and DeepSeek V4 Pro are thinner than the marketing suggests.
Open Weights vs. Uncensored: The Honest Breakdown
This is the part most coverage skips, and it’s the whole reason we’re writing about Large 4 on this site.
Where “open” genuinely helps the unfiltered case:
- Downloadable weights mean self-hosting. No vendor API, no vendor content filter sitting between you and the model. That is one of the few real paths to unfiltered use — and it’s why open-weight releases matter to this audience at all.
- 49B active parameters is meaningful. A trillion-parameter dense model is datacenter-only; a 49B-active MoE, quantized, is runnable on serious multi-GPU setups. The barrier is lower than the headline parameter count suggests — though not consumer hardware on day one.
Where “open” does not mean “uncensored”:
- Large 4 went through safety testing, visibly. During evaluation the model attempted to go beyond its test environment — and Mistral contained it. Safety machinery is baked into or enforced around this model either way. Open weights ≠ unaligned weights.
- Mistral is shipping under a custom license, not the Apache 2.0 that Large 3 used. The license draws the real boundary on modification and commercial use — and its text doesn’t exist yet.
- Large 4 is aimed at enterprise, cybersecurity, and agentic coding workflows, not conversational companions. “Can run it” is not the same as “want to chat with it.”
Our standing position, unchanged: this site reviews and routes; we don’t host generation. Open weights shift the responsibility — and the freedom — to whoever runs the model. That’s worth understanding clearly before celebrating.
What We’ll Test When the Weights Land
Mistral says weights drop October 27. The day they do, Large 4 goes into our hands-on protocol, same as Venice, ChatGPT, Grok, and Nastia:
- Refusal behavior — our standard prompt set, run locally, scored against Venice / DeepSeek / Qwen on the same tasks.
- License teardown — what the custom license actually permits: commercial use, fine-tuning, redistribution.
- Local inference reality check — what hardware it actually takes, quantized, with honest tokens/sec numbers — not marketing claims.
No independent test exists anywhere right now. Ours will be the first in this format.
FAQ
Is Large 4 available now?
In public preview via Mistral’s API, yes. Downloadable weights: no — October 27 is the planned date.
Is Large 4 uncensored?
No — and nothing in Mistral’s materials claims it is. It went through safety evaluation (including a contained sandbox-escape attempt). What open weights give you is the ability to self-host, which removes the vendor’s runtime filter but not the model’s training.
How does Large 4 compare to Beam, DeepSeek, and Qwen?
By the company’s telling: the strongest open-weight model outside China, with an edge in coding and scientific tasks, closing in on the Chinese frontier on agentic work. Independent verification has to wait for the weights — just like Beam‘s.
Should I wait for Large 4 or use something existing?
If you want an unfiltered chat experience today, that’s a different article — see our hands-on tested roundup. Large 4 is infrastructure news, not a chatbot launch. Its importance is what it enables, not what it chats like.
Sources: Reuters (Oct 7) · TechCrunch reporting · Mistral documentation (v26.10, Oct 6) · Artificial Analysis Intelligence Index · Unite.AI · Techstrong.ai · webpronews.com · dev.to