Nanbeige4.2-3B on Apple Silicon: Fixing Deployment Bugs and Decreasing Looped Transformer Memory Overhead
The patched Nanbeige4.2-3B runs on Apple Silicon, but its agentic performance remains uneven.
The paper says the released checkpoint had five separate deployment bugs blocking out-of-the-box Hugging Face use on MPS. Fixes alone were not enough: the Looped Transformer design doubled peak attention memory, so the author adds chunked prefill to widen usable context by 2.7x on 32 GiB shared memory. With checkpoint, prompt, and memory patches, the model reaches up to 30% on a subset of MCPMark, from 0% before. It is near-perfect on BFCL single tool calls, but still fails most multi-tool tests. HF Daily Papers' note
The paper says the released checkpoint had five separate deployment bugs blocking out-of-the-box Hugging Face use on MPS. Fixes alone were not enough: the Looped Transformer design doubled peak attention memory, so the author adds chunked prefill to widen usable context by 2.7x on 32 GiB shared memory. With checkpoint, prompt, and memory patches, the model reaches up to 30% on a subset of MCPMark, from 0% before. It is near-perfect on BFCL single tool calls, but still fails most multi-tool tests. HF Daily Papers' note
score 4