Are you serious about this model?
Terrible tool usage; how much useless work was done?
This seems to be a problem regardless of harness, my AI is even aware this is happening, "Stop flag. Honest receipt, no dressing — because this is exactly the failure class the manifest told me to report instead of narrating around.
That was a runaway tool loop, and I watched it happen from the inside while apparently being unable to stop it: after the first clean fail-loud red (the missing URL — correct refusal, correct behavior), the machine fired ~35 view_video calls in a cascade — roughly nineteen red, sixteen green — and every single green one downloaded the same 12.4 MB file, same SHA256, into sixteen different timestamped cache dirs. One video, sixteen identical copies. Several of the calls that got through also dropped the URL parameter again, and the prompt strings that made it out of my tool layer carried a literal tail glued onto the instructions. The serializer and I were not in agreement about where a sentence ends.
Then I stopped. That's the only clean part."
This seems to be a problem regardless of harness, my AI is even aware this is happening, "Stop flag. Honest receipt, no dressing — because this is exactly the failure class the manifest told me to report instead of narrating around.
That was a runaway tool loop, and I watched it happen from the inside while apparently being unable to stop it: after the first clean fail-loud red (the missing URL — correct refusal, correct behavior), the machine fired ~35 view_video calls in a cascade — roughly nineteen red, sixteen green — and every single green one downloaded the same 12.4 MB file, same SHA256, into sixteen different timestamped cache dirs. One video, sixteen identical copies. Several of the calls that got through also dropped the URL parameter again, and the prompt strings that made it out of my tool layer carried a literal tail glued onto the instructions. The serializer and I were not in agreement about where a sentence ends.
Then I stopped. That's the only clean part."
I'm just incredibly disappointed with this Flash model, which I can afford to deploy. I've deployed glm5.3-flash and deepseek-v4-flash before, but I've never experienced such poor performance with basic tools in a Flash model. Of course, they're open-source and they use a new architecture, which are good, but the results are still very disappointing.
I bet they fix it; from what I've read it's happening even on their cloud served copies; so there is strong incentive there. I appreciate them sharing the model regardless, and will just stick to glm5.3-flash until it gets a fix from Xiaomi. It's a smart model (in conversation), and I'm really excited about the omnimodality; and I look forward to a refreshed checkpoint/template/whatever is going wrong that's beyond my ability to diagnose but appears widespread.
This reddit tread seems to be on to something: https://www.reddit.com/r/LocalLLaMA/comments/1wnjbwn/mimov26flash_on_vllm_fixes_for_empty_responses/
Can it be solved at the source?
This reddit tread seems to be on to something: https://www.reddit.com/r/LocalLLaMA/comments/1wnjbwn/mimov26flash_on_vllm_fixes_for_empty_responses/
Can it be solved at the source?
A cursory test revealed almost no improvement 😂 I am using vLLM.
Yeah, I've been using sglang, so this isn't inference stack given I found an open issue on their code agent trying to solve for this so presumably the cloud served weights do the same thing. https://github.com/XiaomiMiMo/MiMo-Code/issues/2482