The top open-weight model is a Chinese one, again

Qwen3.8-Max leads the August open-weight ranking at 79, ahead of dots3-note Preview at 68.8 and Ornith-1.5-397B at 68.5. A ten-point gap at the top of a leaderboard usually means the second tier has not caught up rather than that the leader is safe.

Qwen3.8-Max leads the August 2026 open-weight ranking at 79, with dots3-note Preview at 68.8 and Ornith-1.5-397B at 68.5 behind it. Separately, Moonshot's Kimi K3 — a 2.8-trillion-parameter mixture-of-experts model — is reported at fourth globally across all models, open and closed.

The gap is the finding

Ten points between first and second is unusual on a mature leaderboard. Bunching at the top normally indicates a field converging; a clear gap indicates either a genuine capability lead or a benchmark that suits one model's training distribution. Both are worth knowing and only one is worth paying for, and the ranking alone cannot tell you which it is.

Where these models come from

The names at the top are predominantly Chinese labs. That is not a new observation in 2026, but it has become sufficiently routine that it now goes unremarked, which is its own kind of shift — a year ago it was the headline.

It also puts the licence question in sharper relief. An audit this month found more than a third of "open" models carry real commercial conditions, including regional exclusions — a term with rather more force when the licensor and the deployer sit in different jurisdictions.

What a score of 79 is not

It is not a deployment recommendation. A leaderboard position measures capability on a fixed set; it says nothing about the licence, the provenance of the training data, the availability of the weights in your region, or whether anyone will answer when the model does something strange in production.

BenchLM — Open-Source LLM Leaderboard 2026: 101 Models Ranked → · Kilo — Best Open Source AI Models for Coding (2026) → · Local AI Zone — Latest AI Developments: August 2026 Update → · Digital Applied — We Read the Licences on 2026 Open-Weight Models →