Mistral Le Chonk: open-weight model vs. China
The trillion-parameter release is pitched as Europe's answer to Chinese open weights, yet Reflection claims the same crown for its Beam model.
In short
Mistral claims its new trillion-parameter model is the strongest open-weight release outside China, but Reflection stakes the same claim within days for Beam, which was trained on 10,500 GPUs.
At a glance
- Mistral calls its new trillion-parameter model the strongest open-weight system outside China.
- Reflection's Beam: 501 billion parameters, 23 billion active per token, Apache 2.0, weights in October 2026.
- Beam was trained on 10,500 Nvidia GB300 GPUs over more than four weeks and more than 80 million RL rollouts.
- Unverified: Le Chonk's benchmarks, license and release date — the Wired report could not be retrieved.
Mistral says its new trillion-parameter model is the best open-weight system available outside China. That ranking cannot be checked independently right now. Within days, US lab Reflection staked the same claim for its own model, Beam.
What Mistral is actually claiming
The French lab frames the trillion-parameter release as proof that it is still in the race for frontier AI. Beyond that framing and the parameter scale, nothing about Le Chonk could be confirmed for this article. Its benchmark scores, license terms and availability remain open questions.
Reflection's competing claim
Coverage dated October 6, 2026 puts Beam at 501 billion total parameters, with 23 billion active per token, released under Apache 2.0. The weights are due later in October 2026 after final safety testing, and an early build is already with selected users. Training ran on 10,500 Nvidia GB300 GPUs for more than four weeks and more than 80 million reinforcement-learning rollouts.
Reflection says Beam matches GLM 5.2 while spending three to four times less compute, and reports 80.9 on SWE Bench Verified.
| Benchmark | Beam | GLM 5.2 |
|---|---|---|
| DeepSWE v1.1 | 44.4 | 44.0 |
| SWE Bench Pro v1 | 65.5 | 62.1 |
| Terminal Bench v2.1 | 80.1 | 81.0 |
Why “outside China” carries the weight
Both claims hinge on the same carve-out, and that carve-out does real work. On the same scoreboard, Kimi K3 posts 68.0 on DeepSWE v1.1 and 88.3 on Terminal Bench v2.1, ahead of Beam, while Qwen 3.8 Max reaches 51.0 and 86.6. The open-weight frontier still sits in China; what Western labs are contesting is the tier below it.
The lab behind Beam
Reflection was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both previously at Google DeepMind. A 130 million dollar seed round in March 2025 was followed in October 2025 by a 2 billion dollar Series A at an 8 billion dollar valuation, with Nvidia among the backers. Beam exposes a tunable reasoning depth, and reportedly picked up web browsing on its own even though browsing tasks were absent from the reinforcement-learning mix.
What we could not verify
The Wired report on Le Chonk would not load when we fetched it for this piece. Every Beam figure here traces back to the linked source; for Le Chonk we have the claim and the parameter scale, nothing measurable. A head-to-head comparison is therefore not possible yet.
FAQ
What is Mistral's Le Chonk?
Le Chonk is Mistral's new trillion-parameter model, which the lab describes as the strongest open-weight system outside China. Its benchmarks and license are not yet confirmed.
Is Le Chonk better than Reflection's Beam?
There is no basis to say so yet. Beam has published numbers, including 80.9 on SWE Bench Verified; Le Chonk has none so far.
Why do both labs say “outside China”?
Because Chinese open-weight models still lead. Kimi K3 scores 68.0 on DeepSWE v1.1 against 44.4 for Beam.
