Researchers identified and fixed five deployment bugs that prevented the Nanbeige4.2-3B language model from running correctly on Apple Silicon, including a silently zeroed RoPE buffer that was stripping out positional information. They introduced chunked prefilling to address the doubled attention-memory overhead of the model’s Looped Transformer architecture, extending usable context width by 2.7 times. After also fixing system-prompt handling and MPS-specific memory bugs, the patched model completed 30% of MCPMark agentic tasks with near-perfect accuracy on single tool calls, though multi-tool scenarios remained problematic.
