🫨绝了啊！有人把完整CPU + 另一个LLM塞进Qwen3.5-4B权重里，还跑出了故事！
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.
🫨绝了啊！有人把完整CPU + 另一个LLM塞进Qwen3.5-4B权重里，还跑出了故事！ 有人用Qwen3.5-4B做了件疯狂的事：直接把完整32位RISC-V CPU编译进模型权重，再用它跑出另一个LLM（Karpathy的stories260K）。 全程在模型自己的forward pass里完成，零外部运行时、零工具调用，数学精度直接从55%干到97.5%+，还是bit-exact的。 核心亮点： 🔹CPU硬核嵌入：11层编译出来的RV32IM（+Zba+大数单元）直接塞在Qwen前面，权重全是编译器算出来的，没训练。 🔹RAM就是模型自己：用Gated DeltaNet状态当精确可读写的RAM，寄存器、PC、栈全在里面。 🔹从CPU进化到NPU：QwenPU进一步编译了向量单元（矩阵乘、softmax、注意力），把stories260K当成C程序跑在Qwen内部，权重也存在DeltaNet里。 🔹效果炸裂：DeepMind大数乘法55% → 97.5%，部分甚至超过带Python工具的基座；全链路bit-exact，230k+步零误差。 🔹完全开源：代码MIT，权重Apache-2.0，可直接加载。 快速上手步骤： 1️⃣克隆代码 git clone https://github.com/divagr18/qwenputer（CPU版）或 https://g
Posted by SuSu_酥酥👅 (25.5k followers) 1 h ago · 14 likes · 730 views · view the original post on X. Kept by the Dev Radar as AI dev tools. Tools mentioned: qwenputer, qwenpu, chipformer.rv.chat, storyrun.py, qwen3.5-4b-qwenputer.
More dev work like this
- Your AI just got a sales team. — @theayush
- My most fun way of learning nowadays is asking AI to explain things to me. — @Chris_Wozniczek
- Meet DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-IQ2_XXS-GGUF: a mouthful, but a beast!… — @HuggingModels
- Supervision, Open-source a library that turns any detection model into production-ready… — @0x0SojalSec
- Chat with your documents — without building a RAG app from scratch. — @DanKornas
- MLX-Serve 26.11.1 - Released, change log & download here:… — @ddalcu
- 10 GitHub projects every MCP developer should explore 🔌 — @PrakashS720
- ❗️ Andrej Karpathy says Claude built him HackerTube, a YouTube that only recommends… — @IntCyberDigest
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. 31.4k posts from 5k X accounts over the last 21 days, 9k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-01 22:12 UTC. Full method.