Nanbeige4.2-3B Model Optimized for Apple Silicon Performance

A recent study addressed critical deployment issues for the Nanbeige4.2-3B model, a 3-billion parameter agentic model utilizing a Looped Transformer architecture, when running on Apple Silicon. The research team pinpointed five distinct bugs, including a silently zeroed RoPE buffer and outdated API calls, which previously hindered the model's out-of-the-box functionality via Hugging Face transformers. Beyond bug fixes, the team tackled the model's memory inefficiency, a consequence of its layer-reuse strategy that effectively doubles peak attention memory. They introduced a chunked-prefill strategy, which successfully reduced memory overhead and expanded the allowable context width by 2.7 times on systems with 32 GiB of shared memory. With these improvements, the debugged model now reliably performs on standard benchmarks. It completes up to 30% of real agentic tasks, a significant improvement from its original 0% success rate, and achieves near-perfect results on single tool calls, though multi-tool tests still pose challenges. The patched checkpoint and related tools are publicly available.