DeepSeek V3 (MLA, DeepSeekMoE, aux-loss-free balancing, MTP) in ~200 lines of pure JAX, trained on a MacBook Air.
A Chinese-first small language model built from scratch: open 60B-token dataset, tokenizer, mHC + Engram + MoE model, pretraining through RL, all in PyTorch. 从零做中文小模型,数据和代码全部开源
Public code release for the paper "The Surprising Effectiveness of Deleting Weights in LLM Post-Training"
Fine-tuning a 0.5B model on my own WhatsApp replies, and measuring what it really learned
Agent Skill for fine-tuning an agentic model on your own Claude Code sessions with Tinker. Works in Claude Code, Codex and OpenCode.
Speculative Self-Distillation enables Efficient Knowledge Internalization (NeurIPS 2026)
A from-scratch, dependency-free GPT implementation in pure NumPy: byte-level tokenizer, multi-head self-attention, Adam trainer, and top-k sampler.
Frontier-level agent capabilities with nothing more than high-quality data and a simple training recipe: 2 SFT + RL, no mid-training/DPO/OPD. Full recipe released.
From the First Token: bilingual book, guided exercises, runnable examples and teaching data
零基础入门大模型(LLM)的基础教程:从零手写 GPT-2,一行行看懂大模型是怎么工作的 | Build GPT-2 from scratch in PyTorch, step by step
Soft-MTP: Multi-Token Prediction Done Right. KL gradient recovery for faster speculative decoding and better language model training.
手写小模型,看懂大模型。20 讲图解与代码,浏览器直接训练,无需 GPU。|Build a tiny LLM from scratch with 20 illustrated lessons and browser training. No GPU required.
ChronoReplay: One Answer Interface Is Not Enough for Streaming Spatial Intelligence