Request too large (max 32MB)

[cnsiva]

I am using claude code to connect my local oMlx instance. It is a medium level code base and getting this error very often. I need to reload the model or restart the omlx server when I get this error. I am using Qwen3.6-35B-A3B model.
Here is the model settings

[jundot]

Thanks for the report. I traced this to oMLX's prefill memory guard, not an actual 32MB HTTP request-body limit.

The server was returning HTTP 413 for PrefillMemoryExceededError, which can make Anthropic-compatible clients like Claude Code surface it as a generic "Request too large (max 32MB)" error. I'm changing this path to return HTTP 400 with code / omlx_code: "prefill_memory_exceeded" plus the estimated and limit bytes, so the client can show the real cause.

If this happens before the fix is available, the practical workaround is to set Memory Guard to aggressive, raise the custom memory guard ceiling, free memory, or compact/reduce the context.

[cnsiva]

@jundot Is this addressed in the latest update?