Mistral Ships 'Le Chonk': Trillion-Param Open Weights
By @sharedot · · 8 pages
- AI Frontier
- Open Weight Models
- Mistral
Mistral's Large 4, codenamed Le Chonk, is a trillion-parameter open-weight model with 49B active params, with weights due Oct. 27.
The meme became a real trillion-parameter model
In June, a fictional Mistral model called 'Le Chaton Fat' went viral across X and Reddit with fake benchmark charts claiming specs like '1,000 meows per second.'
Now the joke has landed: Mistral is launching a public preview of Mistral Large 4 (ML4), a one-trillion-parameter multimodal model internally codenamed 'Le Chonk,' VentureBeat reports.
The company says ML4 pushes it back toward the global open-weight frontier, targeting coding, cybersecurity, finance, manufacturing and visual grounding.
When asked whether ML4 was the model the meme anticipated, chief scientist Guillaume Lample said Mistral had enjoyed the meme and suggested ML4 could be viewed as an initial version of the idea, with still larger models to come.
Sparse architecture: 1 trillion total, 49 billion active
ML4 uses a sparse mixture-of-experts style design: although it contains one trillion parameters in total, only 49 billion are active during inference, per VentureBeat.
The model was trained from scratch over roughly two months on 4,000 Nvidia Grace Blackwell GPUs in Mistral's own European data centers, and across more than 160 languages including every official EU language.
As a comparison point, Mistral says its previous Large 3 model — 675 billion total parameters with 41 billion active — was trained on 3,000 Nvidia H200 GPUs. The new model accepts multimodal inputs but still produces text-only output, Lample confirmed in an interview with VentureBeat.
A three-week preview before open weights on Oct. 27
Mistral plans to publish the model weights on October 27 after roughly three weeks of testing with developers, cybersecurity leaders and government authorities, according to VentureBeat.
The weights will ship under a custom Mistral license, and no API pricing was provided in the materials VentureBeat reviewed. The preview period also gives Mistral time to continue reinforcement learning and tune the final checkpoint before release.
WIRED adds that the model is freely available in preview now, with the final version to follow by the end of the month, and that Mistral positions Le Chonk as by far the most capable open-weight model developed outside China.
Benchmarks are strong — but the picture is complicated
VentureBeat reports that Mistral's supplied charts show ML4 at 62% on DeepSWE v1.1, ahead of Reflection AI's Beam (44%), Qwen 3.8 Max (51%), DeepSeek V4 Pro 0813 (57%) and GLM-5.3 (61%) — but the live DeepSWE leaderboard, selecting best published configurations, puts GLM-5.3 and Kimi K3 near 69% and GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 around 74%.
On Harvey's Legal Agent Benchmark, Mistral claims a 15% task-pass rate, ahead of the open-weight rivals in vals.ai's leaderboard if produced under the same methodology.
VentureBeat could not independently locate published results for some newer-model scores in Mistral's Finch and visual-grounding charts, and ML4 does not yet appear in Artificial Analysis' public evaluations.
Why it matters: sovereignty and the open-weight bet
Mistral is pitching ML4 as a foundation enterprises and governments can run and customize themselves, including on sovereign infrastructure with zero-data-retention options, VentureBeat reports.
The cybersecurity angle is strategically charged: Mistral argues that open weights let security teams run code scanning and defensive testing without depending on a closed provider's changing moderation policy.
Lample told WIRED that 'if you use a closed model, there is no guarantee it will still be there tomorrow,' noting ownership matters even for US companies.
The release follows a September €3 billion Series D at a post-money valuation above €21 billion — which Reuters reportedly valued at roughly $24 billion — the largest equity raise by a European tech company, per VentureBeat.
What comes next after Oct. 27
The practical question for builders is whether ML4's final checkpoint holds up once independent evaluators can test it, since the ranking claim remains provisional until outsiders can run the released weights, VentureBeat notes.
Mistral says its science team has scaled from three researchers to roughly 300, and Lample expects capabilities to improve as reinforcement learning concludes and training capacity expands.
The company's broader bet is that model weights become commoditized while value shifts to the surrounding stack — deployment, infrastructure, customization and engineering support — serving more than 125 enterprises including Airbus, ASML and HSBC, according to VentureBeat.
After October 27, the 'Le Chonk' nickname will matter less than what developers can reproduce on their own hardware.