Loss of Final Tokens After Tool Call When Using kimi-k2 Model with Streaming Output

[gscfwid]

Question

Has anyone encountered an issue where, when using the kimi-k2 model, tokens at the end of the output disappear? I observed this after a tool call, but I’m not sure if it can also happen without a tool call, as I haven’t tested that case. I’m using streaming output (‎agent.iter() + ‎node.stream), and not only is the streaming output incomplete—the ‎final_output is also missing the last few tokens.

Additional Context

Pydantic-AI version: 0.7.4
Python version: 3.12

[DouweM]

@gscfwid Can you please share how you're running kimi-k2 (e.g. ollama, some cloud provider) and which model and provider classes? The code you're using an the full trace of events generated by iter would also be helpful.

[github-actions]

This issue is stale, and will be closed in 3 days if no reply is received.

[github-actions]

Closing this issue as it has been inactive for 10 days.