Cohere Releases Investigation Traces and Scripts for Qwen Model Behavior
Cohere has released investigation traces, raw tokens, and a minimal reproduction script on Hugging Face to document a rare behavior observed in the Qwen model. Developers can now use these artifacts, including specific prompts with prefilling, to test and study the behavior independently. This release underscores the critical importance of open-weight models for AI safety, transparency, and security research. By having access to model weights and detailed traces, the broader community can investigate unusual behaviors that might otherwise remain hidden in proprietary systems. The investigated behavior was initially spotted by developers running Qwen3.8 locally on their own hardware, manifesting as strange links. Cohere noted that while the behavior can occur spontaneously, it is rare, but can be reliably reproduced using a specific prompt combined with prefilling.
## BACKGROUND
LLM prompt prefilling is the initial phase of inference where a model processes an entire input prompt in a single parallel pass to build up the Key-Value (KV) cache. Open-weight models allow researchers to inspect exact weight artifacts and internal states, enabling deep security analysis that is impossible with closed, API-only models.