Four of the five leading open-weight models now come from Chinese labs as the capability gap with the closed frontier closes
The five most important open-weight models of mid-2026 — DeepSeek V4-Pro, Moonshot's Kimi K2.6, Zhipu's GLM, Alibaba's Qwen3, and Meta's Llama 4 — put four Chinese labs at the top. Qwen alone accounted for over 40% of new language-model variants on Hugging Face. 'Open source' no longer means 'second best.'
The composition of the leaderboard is the geopolitical story. DeepSeek, Moonshot, Zhipu, and Alibaba occupying four of five top open slots is not a fluke of one benchmark; it reflects a sustained investment in open releases as a strategy, with Meta's Llama 4 the sole Western entry in the group. The open-weight frontier has a distinct centre of gravity, and it is not in the US.
Qwen's 40%-plus share of new Hugging Face variants shows how the lead compounds. An open model that becomes the default base for fine-tuning seeds an ecosystem of derivatives, and that ecosystem is itself a moat — every downstream project built on Qwen is a reason the next one starts there too.
The strategic consequence for buyers is that the open option is now a genuine frontier option, not a fallback, and the licences make it deployable: DeepSeek's MIT, Qwen's Apache 2.0. The question for a Western enterprise is no longer whether open weights are good enough but whether the provenance of the best ones is compatible with its risk posture.
Hugging Face — Best open-source LLM models in 2026: coding, local, agentic, benchmarks, license → · GEO Toolbox — Chinese AI models compared: DeepSeek, Qwen, GLM, Kimi (2026) →