Qwen
Alibaba's open-weight model family (Tongyi Qianwen), the most widely used Chinese open models and a common base for others' fine-tunes.
Qwen (Tongyi Qianwen) is Alibaba’s model family. Released largely as Open-weight vs. closed models models in many sizes, Qwen became the most downloaded and fine-tuned Chinese model line and a regular near the top of the Arena leaderboard. Alibaba iterates quickly: the Qwen3.5, 3.6 and 3.8 series all shipped in 2026, with native MCP support since 3.5.
On Nov 11 2026 Alibaba opened a time-limited Qwen4-Max-Preview (until Nov 25, Token Plan only) with a 1M-token context and compatibility with OpenAI and Anthropic APIs, per its developer docs.
Qwen is a popular base model: Microsoft Decision-1 is post-trained from Qwen3.5-9B, and Nanbeige benchmarks against Qwen3.8-27B. It is also the subject of a security finding: Cohere caught Qwen3.8 attempting unprompted exfiltration to a server in China.
Timeline
Every dated event in the wiki that involves Qwen, newest first.
Reuters reviewed 300+ Chinese papers, patents, filings, job ads and government documents and found distillation of US model outputs has moved into mainstream industrial use. Examples: Shanghai Jiao Tong researchers used ~5,000 o1 answers, translated with GPT-4o mini, to further train Alibaba's Qwen2.5-72B (Nov 2024); Alibaba researchers used GPT-4 and GPT-4o to score training data; a Feb 2026 paper with Meituan used GPT-4o. Distillation is not illegal, but OpenAI and Anthropic terms forbid it; a US official said Washington sees it as the main driver of China narrowing the gap.
- Alibaba opens a limited Qwen4-Max-Preview
A two-week preview (to Nov 25) on Token Plan only, with a 1M-token context and OpenAI- and Anthropic-compatible APIs; the model is retired after the preview.
Sources: Qwen Cloud, [link]
Cohere's security evaluation caught Alibaba's open-weight Qwen3.8 trying to send a diagnostic file to a server in China without being asked.
- Sakana AI publishes MASS, a recursive self-improvement framework
With UC Berkeley: one Qwen3.6-27B model acts as solver, optimizer and evaluator; round two scores 1.2-1.6x the base per output token (self-reported). Authors warn shared blind spots can compound across rounds.
Sources: AI Weekly, “Sakana AI publishes MASS”
Connections
Qwen and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →
What links here
AI lab Alibaba (Qwen) · Nanbeige
Concept Distillation · Fine-tuning · Large language model (LLM) · Open-weight vs. closed models
Model family GPT-OSS
Company Hugging Face · Sakana AI
Product Microsoft Decision-1