On AI.
关注 AI，也动手做 AI —— 训练、推理、评测，以及开源生态。
LLM training · inference engines · model evaluation · open-source AI
Run the native 284B-A13B DeepSeek-V4-Flash-0731 checkpoint on one laptop CPU. The pure C engine streams cold MoE experts from disk, has a tested 8 GB RAM path, needs no GPU or Python, and reaches 1.12 token/s in its best documented prompt-lookup workload. It supports terminal chat and a resident local OpenAI-compatible API with live token streaming, parallel function calls, and matched tool-result replay.
Run Qwen3.8-27B locally on one laptop CPU with a native C engine, direct GGUF loading, an 8 GB RAM tested path and measured generation up to 2.52 token/s. Chat in the terminal or connect local apps through its resident OpenAI-compatible API with parallel function calls and tool-result replay. Runtime speedups preserve byte-identical full logits against the native baseline for the same GGUF.
- Training: RL-scaling post-training frameworks for LLMs, plus training experiments that fit on a laptop CPU — with benchmark evidence
- Inference: native engines (pure C) that bring large open models to ordinary hardware, with reproducible performance and correctness evidence
- Evaluation: probing what models truly believe — profiling the ideology of any foundation model
- Open-source AI: curating and maintaining resources across LLM inference, local AI, small language models, Chinese LLMs and AI developer tools