Mistral releases Mistral Large 4 (“le Chonk”)
1T-parameter open-weight MoE, 1M context, public preview
Confirmed
Importance: major (4 of 5)The takeaway
On Oct 6, 2026 Mistral AI released Mistral Large 4, a granular mixture-of-experts model with ~1.05T total and 49B active parameters, a 1M-token context window and vision, in public preview on Mistral Studio; open weights are promised by the end of October 2026.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- Major (4 of 5)
- Last verified
- 29 October 2026
Your AI and this story
- GPT-6.2 Astra128 days after its cutoff
- Claude Opus 5.598 days after its cutoff
- Gemini 4 Hydrogen189 days after its cutoff
- Grok 4.7128 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 98 days before it.
Key facts
- Architecture: granular MoE, 49B active / 1.05T total parameters, 1.6B vision encoder (docs)
- Context window: 1M tokens (docs); API id mistral-large-4-0 / mistral-large-4
- Trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters (blog; Techmeme/The Deep View said ~4,000)
- Availability: public preview via Mistral Studio from Oct 6; weights to be released by end of October 2026
- Pricing differs between sources: docs list $0.68 input / $0.07 cached / $2.09 output per 1M tokens; the blog lists $1.36 / $4.18 (docs figure may be a preview price - unverified)
- Mistral-reported benchmarks: Cybench 93%, DeepSWE v1.1 61.7%, Terminal Bench 4.0 28.3%, AutomationBench 59.9%, Dense 200 42% (vs GPT-6-Astra 41%); top 5 on the AA Cyber Index
- Hacker News: the blog thread reached ~1,666 points and ~947 comments by Oct 7 (627 points for the docs page on launch day). Commenters noted a 50% launch discount to compete with DeepSeek V4.1 Flash, a likely source of the two price figures (unconfirmed)
- Reach: Mistral’s launch post on X had ~4.04M views and 37k likes by Oct 7; NVIDIA AI’s post on the training run (~4,000 Grace Blackwell superchips) ~250k views
Show 11 more
- Guillaume Lample (X thread, ~197k and ~487k views on Oct 7; the second tweet, on the 3,800-GPU Bruyères-le-Châtel cluster, had ~1.29M views by Oct 8; NVIDIA AI Infra’s training-run post ~286k): the preview is a checkpoint of an RL run ‘still in flight’ that ‘shows no sign of saturation’; a final version will ship with the weights before the end of October; trained on 3,800 Grace Blackwell GPUs in Mistral’s Bruyères-le-Châtel cluster, funded by its Series B, with Series C and D clusters ‘coming online soon’
- Arthur Mensch: ‘Trained and served on our own compute, and RL shows no sign of saturation’; Mistral researcher Albert Q. Jiang said ‘an even chonkier boi is munching on trillions of tokens’, meaning a larger model is in training
- Launch post: ‘Working with cybersecurity partners privately’ and deployable from Europe via Mistral Cloud; Hugging Face placeholder repo mistralai/Mistral-Large-4.0-1T05-A52B (linked by Julien Chaumond)
- Early tests: Simon Willison found only reasoning ‘none’ or ‘high’ settings and called its pelican SVG the best from any Mistral model; Arena added it to Agent Arena and Code Arena (scores pending). HN users asked whether it was OpenRouter’s stealth model ‘space-bunny-alpha’ (unconfirmed)
- Artificial Analysis (Oct 6–7): Intelligence Index 38, ‘comparable to GPT-6 Luna (max, 38) and DeepSeek V4.1 Flash (max, 39)’, ‘the most intelligent model from outside the US and China’; Cyber Index 50 (level with GLM-5.3-Flash); GDP.pdf 19% (+18 pts vs Mistral Large 3); AA lists a 512k context in its test setup; $1.36/$4.18 per 1M standard, 50% off ($0.68/$2.09) for the first two weeks, which resolves the pricing discrepancy; ~$1.13 per task at standard price
- VentureBeat: custom Mistral license; open weights dated Oct 27; 160+ languages; other Mistral benchmarks Harvey Legal Agent 15%, FinWorkBench 67%, DIOR-RSVG 73%; Arthur Mensch joked about the ‘le chaton fat’ meme (‘It’s actually le gros chaton’)
- Hugging Face ‘Upcoming release’ page for mistralai/Mistral-Large-4.0-1T05-A52B (linked by HF co-founder and CTO Julien Chaumond; read 7 Oct): ‘Expected release October 31, 2026’ (VentureBeat said Oct 27); ‘49 billion active parameters per token (52 billion including embeddings and output layers)’, which explains ‘A52B’ in the name; 579 people waiting
- Same week setback: Berlin-based search engine Ecosia (used by Germany’s federal environment ministry and the UK NHS) told POLITICO (CEO Christian Kroll interview, published Oct 6–8) it is dropping Mistral, its AI supplier only since May 2026 (when it left OpenAI), for open-weight models including Qwen, GLM and Kimi served by the German platform Melious on EU servers. Kroll: Mistral is ‘a year behind’, ‘We were simply too large a customer for Mistral’, and the switch ‘roughly cut our costs in half while improving quality’ (ppc.land, Cybernews, Trending Topics). Lample, asked at an Oct 5 press conference, said: ‘I would encourage them to test our new model as soon as possible’ (per ppc.land)
- Tooling: Simon Willison’s llm-mistral 0.16 (Oct 6) adds support for Mistral reasoning models such as Large 4
- Weights released Oct 29, 2026 (see Mistral releases the weights of an improved Mistral Large 4 (“Le Chonk”), a 1T-parameter MoE, under a modified MIT licence): a further-trained checkpoint (Terminal-Bench 4.0 41% vs 28% in preview, Mistral-reported) as mistralai/Mistral-Large-4-1T-A52B (FP8, NVFP4) under the Mistral Ecosystem Open-Weight licence (modified MIT; model-as-a-service firms need a commercial licence); API price cut to $0.70/$2.10 per 1M tokens
- Mistral–Nvidia industrial-engineering collaboration (Oct 21, 2026): Mistral says Large 4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral-operated European data centers, repeats that weights are planned for the end of October, and announces public physics-model benchmarks, open datasets and agentic CAE/EDA benchmarks (no names yet) plus a frontier base model with the Nemotron Coalition; Mistral targets 1 GW of compute by 2030
What happened
Mistral AI launched Mistral Large 4, nicknamed “le Chonk”, on October 6, 2026. It is a granular mixture-of-experts model with about 1.05 trillion total and 49 billion active parameters, a 1.6B vision encoder and a 1M-token context window. It launched as a public preview on Mistral Studio, and Mistral promises open weights by the end of October. Mistral pitches it for general agentic work and highlights its cybersecurity scores (Cybench 93%, top 5 on the Artificial Analysis Cyber Index). All benchmark numbers are self-reported.
The license for the weights and the final price are not yet confirmed: the docs and the blog list different prices.
Reactions on X (views at 11:31 UTC on 7 Oct): CEO Arthur Mensch wrote “Trained and served on our own compute, and RL shows no sign of saturation” (~111k). Quoting Lample’s tweet about the Bruyères-le-Châtel cluster, which shows a photo of the Breuillet – Bruyères-le-Châtel railway station, he added “We’re building AGI from this train station” (~130k). Peter Gostev (Arena) contrasted “4,000 GPUs” with Astra’s “100,000” (~162k). Both figures come from earlier claims, not new disclosures: NVIDIA said “nearly 4,000” and Mistral 3,800, while Astra’s “more than 100,000 GPUs” is in OpenAI releases GPT-6 Astra, its first GPT-6 model. Memes also travelled: synthwavedd joked that the one benchmark ML4 leads is “about regulation” (~156k), and stupidtechtakes posted “Mistral using AI 💔” (~53k).
Why it matters
It is the first trillion-parameter open-weight model from a European lab, and it is trained on European compute. It competes directly with Chinese open-weight frontier models.
Sources
21 sources from 13 sites. Numbers match the chips in the text.
21 sources: 8 primary, 6 press, 7 reactions
Primary
- Mistral: Mistral and NVIDIA bring AI to industrial engineering (Oct 21, 2026)mistral.ai, official
- Simon Willison: llm-mistral 0.16simonwillison.net, code
- Artificial Analysis: Mistral Large 4 puts France back on top outside the US and Chinaartificialanalysis.ai, docs
- Mistral AI: Mistral Large 4 “Le Chonk”mistral.ai, official
- Mistral docs: Mistral Large 4 model carddocs.mistral.ai, docs
- Mistral AI on X: Meet Mistral Large 4, aka Le Chonk (~4.0M views)x.com, official
- Guillaume Lample on X: ML4 launch threadx.com, official
- Hugging Face: Mistral-Large-4.0-1T05-A52B (upcoming release page)huggingface.co, docs
Press
- POLITICO Europe: Ecosia drops Mistral for open-source models, including Chinese onespolitico.eu, press
- PPC Land: Ecosia drops Mistral AI for open models, citing a 1-year quality gapppc.land, press
- Ars Technica: Mistral says “Le Chonk” can challenge the best AI modelsarstechnica.com, press
- VentureBeat: Mistral debuts Large 4 ‘Le Chonk’, a 1-trillion-parameter model with open weights plannedventurebeat.com, press
- The Deep View: Why Mistral’s 1T model is a hedge against lock-inthedeepview.com, press
- TechCrunch: Mistral’s new 1T model aims to leapfrog closed and open rivalstechcrunch.com, press
Reactions
- Simon Willison: Mistral Large 4 (HN thread notes)simonwillison.net, discussion
- Hacker News discussion (~1,666 pts)news.ycombinator.com, discussion
- Simon Willison: Introducing Mistral Large 4: Le chonksimonwillison.net, discussion
- Arthur Mensch on X: ‘Trained and served on our own compute, and RL shows no sign of saturation’ (~111k views)x.com, discussion
- Arthur Mensch on X: ‘We’re building AGI from this train station’ (~130k views)x.com, discussion
- Peter Gostev on X: 4,000 GPUs vs Astra’s 100,000 (~162k views)x.com, discussion
- Julien Chaumond (Hugging Face) on X: link to the weights pagex.com, discussion
Also cited in the text
- synthwaveddx.com
- stupidtechtakesx.com
Changes
- Weights released; pointer to the new entry Mistral releases the weights of an improved Mistral Large 4 (“Le Chonk”), a 1T-parameter MoE, under a modified MIT licence
- Added the Oct 21 Mistral–Nvidia post (3,800 GB-series GPUs, weights end of October, CAE/EDA benchmarks)
- Added Ecosia dropping Mistral for open Chinese models (POLITICO via ppc.land), Ars Technica and llm-mistral 0.16 links
- Added X posts (Mensch x2, Peter Gostev, Julien Chaumond, meme reach) and the Hugging Face upcoming-release page (ETA Oct 31, 52B active incl. embeddings)
- Linked the video “Meet Mistral Large 4, aka Le Chonk”
- Added Artificial Analysis scores (Index 38, Cyber 50) and pricing clarification, VentureBeat details, The Deep View and Simon Willison links
Show all 8 changes
- Added X reach and first-hand posts (Mistral, Lample, Mensch, Jiang), HN thread details and early tests
- Filed