🚀 You don't need a cloud Copilot bill for autocomplete on a Mac anymore. Someo… · @QwenAIFans · TwKit

@QwenAIFans

🚀 You don't need a cloud Copilot bill for autocomplete on a Mac anymore. Someone just shipped Qwen2.5-Coder-1.5B as a Core ML build for the Apple Neural Engine — ~1.3GB, offline, plugs into VS Code via ane-fim. On an M3 16GB: 🔹 ~30 tok/s decode 🔹 ~140ms typical editor response (KV reuse) 🔹 57% exact-match line completion (edit sim 0.82) vs the 0.5B build: 43% / 0.72 Caveat: only one Mac tested so far, greedy decode only, fixed 2k context. via Hugging Face 👇 https://huggingface.co/tianharjuno/Qwen2.5-Coder-1.5B-ANE-ctx2048-LUT6

曝光 184 · 评论 1 · 点赞 2 · 书签 1