Something is afoot in the land of Qwen (simonwillison.net)

Article summary

The Qwen team, responsible for developing open weight models, has seen several key members resign, including the lead researcher Junyang Lin. The resignations may be related to a reorganization within Alibaba, where a new researcher was put in charge of Qwen. The Qwen 3.5 models, released recently, have been impressive, and it is unclear how the resignations will affect the team's future work. The departures have sparked concerns about the project's continuity and the potential loss of talent.

Main themes

• Qwen team resignations
• Open weight models
• Alibaba reorganization
• AI model development
• Research team dynamics

What commenters say

• The Qwen 3.5 models are exceptionally good, but their development may be hindered by the resignations of key team members.
• The Qwen models are overfit to specific tasks and struggle with broader capabilities, such as utilizing tools or commands.
• Some commenters have found the Qwen 3.5-35B-A3B model to be more capable than the Qwen3-Coder-Next model, while others disagree.
• The choice of harness or coding agent can significantly impact the performance of the Qwen models, with some finding success with specific setups.
• The behavior of the Qwen models, such as getting stuck in infinite loops, may be intentional or emergent, and some see it as a trade-off for their tenacity in solving problems.
• The use of MoE models and streaming weights from disk can affect performance, but may be necessary for large models, and some argue that it is unusable for real-time responses.
• Different quantization settings and templates can significantly impact the performance of the Qwen models, and some have found success with specific combinations.
• The Qwen models are still behind commercial models, and some commenters are concerned about the potential loss of talent and the project's future continuity.