Mistral Launches Mistral Large 4, a 1T-Parameter Open-Weight Flagship
Mistral announced its flagship open-weight model Mistral Large 4 (codename Le Chonk) on October 6 as a Research Public Preview: a native multimodal MoE with 1T total parameters, 49B active, text and image input, and a 1M-token context window. The API launched the same day, with open weights expected by end of October — the Hugging Face pre-release page (relayed by victormustar) shows an expected October 31 release and lists active parameters as 52B. It is Mistral's largest open-weight flagship and a landmark move for the European open-source camp.
Confirmed
- Specs: 1T total / 49B active parameters (HF page says 52B), native multimodal MoE, text+image input, 1M-token context; visual grounding claimed to beat some closed frontier models.
- Positioning: the company calls it the best open-weight model on aggregate benchmarks in the US/Europe — and outside China; SOTA on cybersecurity defense, manufacturing and finance workloads, optimized for coding and cyber defense.
- Training: co-founder Guillaume Lample disclosed that ML4 was trained end-to-end on 3,800 NVIDIA Grace Blackwell GPUs at a self-built European datacenter in Bruyères-lès-Châtel; NVIDIA called it "frontier-class compute made in Europe." Lample also teased upcoming Series C/D clusters.
- Release cadence: API live at launch, currently Research Public Preview; weights due end of October (expected Oct 31). RL training reportedly continues to bring gains.
- Human evals: Lample said ML4 beats GLM 5.3 on STEM, CAD and finance tasks and ties on agentic coding. LMArena's live channel Le Chonk ran a round of real-world evaluations.
- Reception: per WIRED (relayed by 233C), Mistral positions the model as Europe's answer to Chinese open-weight rivals like DeepSeek and Qwen. Tansu Yegen argued Europe need not beat OpenAI or Google everywhere but should not be excluded from the AI race. Community response leaned memetic ("chonky"). TestingCatalog noted Google and OpenAI model updates the same day.
Unconfirmed
- Community relays note gaps in official benchmark data; independent evaluations are pending, and detailed LMArena results have not been published.
- Claims such as "beats GLM 5.3," "surpasses closed models," and "best outside China" come from official statements without third-party verification. The "competitor Astra used 100k GPUs" comparison cited in relays (steipete, beffjezos) comes from a third-party tweet, not official data. An earlier relayed claim of "second in blind coding tests, behind Opus 5" lacks a primary source and is not adopted.
Why it matters
- A 1T-parameter open-weight model is a landmark for the European open-source camp, directly benchmarked against US and Chinese frontier open models.
- Training a trillion-parameter model end-to-end on just 3,800 Grace Blackwell GPUs inside Europe — if performance claims hold up to third-party verification — challenges the assumption that frontier models require massive compute, and could reshape enterprise model selection.
2026-10-06 ~ 2026-10-08 · 86 related posts
- Episode 1: Mistral to Launch New Flagship Model, Claims Edge Over Chinese Models(2026-10-06, 2 posts)
- Episode 2: Mistral Launches Mistral Large 4, a 1T-Parameter Open-Weight Flagship(2026-10-06, 86 posts)
- Episode 3: Mistral Large 4 Still in RL Training, Expected by Month's End(2026-10-06, 3 posts)
- Episode 4: Mistral Large 4 Review: France Back in Third Place, but Value Questioned(2026-10-06, 11 posts)
- Episode 5: Mistral Large 4 Teased on Hugging Face: 1T-Parameter Open-Weight Model Coming This Month(2026-10-06, 2 posts)
- Episode 6: Mistral Claims Large 4 Among Strongest Cybersecurity AI Models(2026-10-07, 2 posts)
- Episode 7: Mistral launches trillion-parameter multimodal Large 4, open weights coming this month(2026-10-10, 4 posts)
Primary sources
- Mistral launches Le Chonk: 1T-parameter multimodal model, open weights due end of October — MistralAI ·
- Mistral Large 4 leaked on Hugging Face: 1T total / 52B active params, open-source this month — victormustar ·
- Mistral trained ML4 on 3,800 Grace Blackwell GPUs; Series C and D clusters coming soon — GuillaumeLample ·
- Mistral launches 1T-parameter MoE model Large 4, claims best open-weights model in US/Europe — Bam4d · 2026-10-06
- Mistral launches 1T-param Mistral Large 4; users question its benchmark claims — haider1 · 2026-10-06
- ML4 Trained on 3,800 Grace Blackwell GPUs; Mistral's C and D Round Clusters Coming Online — GuillaumeLample · 2026-10-06
- Mistral's ML4 Hits Open-Weight SOTA on Security, Scores 82% on Vulnerability Patching — GuillaumeLample · 2026-10-06
- Mistral ML4 Benchmarks: SOTA on Finance and Legal Workflows, Terminal and Spreadsheet Navigation — GuillaumeLample · 2026-10-06
- Mistral ML4 Human Eval Beats GLM 5.3 on STEM, CAD, Finance; Ties on Agentic Coding — GuillaumeLample · 2026-10-06
- Mistral ships 1T-parameter open-weight ML4, beating GLM 5.3 on human eval, hiring worldwide — GuillaumeLample · 2026-10-06
- Mistral releases 1T-parameter 'Le Chonk', claims best open-weight model outside China — nordicinst · 2026-10-06
- 1T-parameter model fully pre- and post-trained on 3,800 Grace Blackwells in Europe — qtnx_ · 2026-10-06
- [source] Mistral Large 4 leaked on Hugging Face: 1T total / 52B active params, open-source this month — victormustar · 2026-10-06
- Mistral's Sophia Yang: new model beats GLM 5.3 on human evals across domains — sophiamyang · 2026-10-06
- Mistral officially launches Mistral Large 4 in Research Public Preview — MistralAI · 2026-10-06
- Mistral Large 4 debuts: 1.05T parameters with 49B active, weights out October 27 — 赛博禅心 · 2026-10-06
- Mistral releases Large 4 preview: the only Western OSS LLM competitive with Chinese open models on coding — b_roziere · 2026-10-06
- Mistral Large 4 claims wins over GLM 5.3 on DeepSWE and Kimi k3 on Tbench 4 — b_roziere · 2026-10-06
- Mistral Large 4 preview: 1T-param MoE with 49B active, beats GPT-6 Astra on visual grounding — testingcatalog · 2026-10-06
- Mistral announces 1T-parameter Mistral Large 4 'Le Chonk', API live, open weights this month — kyliebytes · 2026-10-06
- Mistral Large 4 preview: 1T params, 49B active, trained on 3,800 Grace Blackwell GPUs — petrusenko_max · 2026-10-06
- Mistral Large 4 trained on nearly 4,000 NVIDIA Grace Blackwell Superchips, backed by €3B Series D — cedric_chee · 2026-10-06
- Mistral claims new open-weight Le Chonk is best outside China — 233C · 2026-10-07
- Mistral Large 4 trained on just 4,000 Grace Blackwell GPUs vs 100,000 for Astra — steipete · 2026-10-07
29 near-duplicate retellings: MistralAI · testingcatalog · reach_vb · qtnx_ · arthurmensch · MaybeLiterally · SIGKITTEN · uwukko · YashasGunderia · GuillaumeLample · sivareddyg · MaziyarPanahi · IsForAt · _AndrewZhao · MistralAI · _AndrewZhao · mervenoyann · Dr_Singularity · marvinvonhagen · _AndrewZhao · mark_k · mark_k · film_girl · Kyrannio · ylecun · emax · GuillaumeLample · damianplayer · _AndrewZhao