AIData brief
China's best open model leads the US by 46 Arena points, down from 72
The best US open model on the board is Nvidia's Nemotron 3 Ultra. Reflection's Beam, billed as the top non-Chinese open model, is not yet scored.
Tec N Trend data desk · 1 min read · 2 open sources
- Best Chinese open model, 2 Oct 2026
- 1491 (mimo-v2.6-pro)
- Best US open model
- 1445
- China lead over US, Sep 2025 to Oct 2026
- 72 to 46 pts
- Best European open model (Mistral Large 3)
- 1428
The Decoder reports that Reflection AI's Beam, a 501B-parameter model (23B active) due under Apache 2.0 later this month, is now the strongest open-weight model built outside China. Reflection says it matches GLM 5.2 on key benchmarks. Beam is not on LMArena yet.
LMArena's text ranking shows what it has to beat. On 2 October the top open-weight model from a Chinese lab was Xiaomi's mimo-v2.6-pro at 1491, ahead of the best US open model (nvidia-nemotron-3-ultra-550b-a55b-nvfp4, 1445) by 46 points and Europe's Mistral Large 3 (1428) by 63. A year earlier the China-US gap was 72.
The gap shrank because US open models jumped (Gemma 4 in March, Nemotron 3 Ultra in July), not because China stalled; its best rose 63 points. A Beam score above about 1445 would make it the top non-Chinese open model on the board.
Data, method and caveats
- LMArena text leaderboard history (overall category), cleaned snapshot table CC BY 4.0 (LMArena data) · captured 2026-10-07
- The Decoder, Reflection Beam report news report (facts only) · captured 2026-10-07
- Reflection Beam is not in the Arena data (last snapshot 2026-10-02, before its release). The news supplies its size and licence only; the chart plots no Beam point.
- "Open weight" uses the leaderboard's licence field, filled from each model's other snapshots where a snapshot had it blank; any licence other than Proprietary counts, including non-commercial ones.
- Lab country is by company headquarters, set by us: China (Alibaba, DeepSeek, Moonshot, Xiaomi, Z.ai and others), US, Europe (Mistral, Stability), Other (Cohere, AI21, TII, Upstage and similar). A model named Trinity with no listed org is counted as US (Arcee).
- Points are the last snapshot of each month from Jan 2024, plus 2 Oct 2026. Arena scores drift as new models and votes arrive, and are only comparable within a snapshot.
- Scores are the default overall text view, not style-controlled or category-specific.
This brief as JSON or Markdown. Reuse the chart with credit to Tec N Trend and the sources above.