[v0.2.20rc1] - Nemotron Cascade still not working in Claude

[HLGRL]

v0.2.20rc1
I tested this version and Nemotron Cascade model exactly this one: "Nemotron-Cascade-2-30B-A3B-8bit" still does not work in Claude. It process prompt in oMLX, but after loading context it return no content in Claude, maybe still some issues with template.

[Far-cloind]

me too.

[AFimin]

Pretty similar with Opencode.

[jundot]

I tested this with both the oQ4e quant and the full precision model of Nemotron Cascade, and confirmed the issue on both. So this isn't a quantization problem either.

After some investigation, it looks like Claude Code's system prompt is around 27k tokens long, and the model struggles to generate meaningful output at that context length. Short prompts (~200 tokens) work correctly. I don't think this is a response parsing issue on the server side since the model itself is generating garbage or almost nothing before hitting EOS.

Still figuring out whether this is a model limitation with long contexts or something the engine could handle better. Need a bit more time to dig into this. Will update here once i have more clarity.