Chinese open-weight thinking models in 2026 from Z.ai, Alibaba, Moonshot and MiniMax
Alongside DeepSeek's V3.2 and V4, Z.ai, Alibaba, Moonshot and MiniMax kept releasing thinking-first open models.
- Date
- 12 February 2026
- Who
- Zhipu (Z.ai), Alibaba Qwen, Moonshot, MiniMax
- Confidence
- Medium (Z.ai dates from its release notes; others via model cards and Wikipedia
- Deep dive
- Reasoning II, from o1 and o3 to DeepSeek-R1 and the labs that replicated them
Tier: Supporting · Significance: 3/5 · Org(s): Zhipu (Z.ai), Alibaba Qwen, Moonshot, MiniMax · Confidence: Medium (Z.ai dates from its release notes; others via model cards and Wikipedia pointers) Alongside DeepSeek's V3.2 and V4, Z.ai, Alibaba, Moonshot and MiniMax kept releasing thinking-first open models. Z.ai's release notes list GLM-4.5 (2025-07-28), GLM-4.6 (2025-09-30), GLM-4.7 (2025-12-22), GLM-5 (2026-02-12), GLM-5.1 (2026-04-07, billed for eight-hour long-horizon tasks), GLM-5.2 (2026-06-16, 1M context), GLM-5.3 (2026-08-18) and GLM-5.3-Flash (2026-08-26) (Z.ai). Moonshot followed K2 Thinking with K2.5 (2026-01, native vision) and K3 and reached a US$35 billion valuation by July 2026 (Wikipedia, a pointer). Alibaba moved from hybrid Qwen3 to a thinking-only open flagship (B08-20). MiniMax-M2 is an open 230B-total, 10B-active interleaved-thinking model under a modified MIT licence (model card; release date not verified here), after M1.
Four patterns stand out. Thinking is the default or only mode in the newest open models. Licences vary and are not all permissive (Apache 2.0 for Qwen3.8-27B per Wikipedia, MIT for DeepSeek V4, and a custom Kimi K3 License that requires Model-as-a-Service operators above US$20 million of revenue in any 12 months to negotiate a separate commercial agreement). Prices rose at the top (K3 at $3 / $15 per million tokens, which Willison calls the most expensive model a Chinese lab had released). And these models are close enough to the frontier to matter operationally, since Hugging Face analysed the July 2026 intrusion with a self-hosted GLM-5.2 after US frontier models refused (B08-49a).
Dates for MiniMax's later models and the reported Hong Kong listings of Z.ai and MiniMax are not verified (Backlog). For the wider picture see B20. Sources: Z.ai release notes · Kimi K3 License · MiniMax-M2 card · Qwen3.8 card · Willison on K3