起初只当"4 块钱两小时训出 64M 模型"是句营销话，直到自己把 MiniMind 的 trainer 目录翻了一遍。
This is a dev post classified by AGK Intelligence as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
起初只当"4 块钱两小时训出 64M 模型"是句营销话，直到自己把 MiniMind 的 trainer 目录翻了一遍。 预训练、SFT、LoRA、DPO、PPO/GRPO、Tool Use、Agentic RL 全在里面，而且都是原生 PyTorch 手写的，没套 trl 或 peft 那层壳。 结构对齐 Qwen3。我在 3090 上粗估，用 mini 数据组合跑完 1 epoch 差不多两个多小时，租卡成本确实压在个位数。对比两年前的 nanoGPT，那更像教学 demo；市面上多数教程又只停在推理和 LoRA 微调。 MiniMind 算少数把后训练整条链路摊开的仓库，想弄明白大模型到底怎么练出来的，从这儿下手比刷视频课实在。 🔗：https://github.com/jingyaogong/minimind
Posted by lumxss (34.7k followers) 59 min ago · 7 likes · 336 views · view the original post on X. Kept by the Dev Radar as AI dev tools. Tools mentioned: minimind.
More dev work like this
- This free Grok Bot alternative shipped 21 feature requests in under 24 hours. INSANE!! — @ProductHunt
- 试了下 Faceplugin 那套开源人脸识别 SDK。Windows 和 Linux 都支持，处理过程全程留在本机。 — @bkdgiffug
- My favorite thing about T3 Code is that it lets your agent delegate tasks to subagents… — @AdamHoltererer
- The world needs two types of document OCR: — @jerryjliu0
- BREAKING: Most "AI + 3D" stacks still force agents through tools built for humans. — @guilleflorvs
- TrenTorch 的路线定得很死：通篇只用 NumPy，不沾 PyTorch，从张量、自动求导一路手写到 attention、量化和 KV cache。骨架取自哈佛… — @bkdgiffug
- 用 ComfyUI 跑视频模型，光把模型、LoRA、采样、分辨率这一串节点连对就要折腾半天，参数还散在画布各个角落。 — @GitHub_Daily
- Qwen 4 built a rocket launch simulator in one HTML file, and it actually reaches orbit 🚀 — @aijoey
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 32k posts from 5k X accounts over the last 21 days, 9.5k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-07 02:32 UTC. Full method.