Watch an LLM think: 3D visualization traces a prompt through all 64 layers of Qwen3.8-27B
gamedevsam · reddit · 2026-10-26
Developer gamedevsam released a 3D visualization tracing an entire inference pass through Qwen3.8-27B, an open-weight dense model with 27.8B parameters. He asked the model "what does a DJ spin," then followed the token through all 64 layers.
The model is depicted as a record on a turntable: each token rides the groove from the outer edge to the label while every unit of every layer works on it. All numbers shown — probabilities, what attention read, how memory changed — were recorded from the model actually running on his PC.
The 10-minute video offers an intuitive answer to why a model small enough to run on a gaming PC can score so well, and makes dense-LLM internals genuinely explorable.
More from Research
- Glorius Group Releases PRISM, a Bidirectional NMR Foundation Model — CatAstro_Piyush · 2026-10-26
- Training recipe shapes neuron interpretability: GPT-2's dropout may skew findings — tszzl · 2026-10-26
- DeepSeek 1017 checkpoint lands on FrontierCS 2 with a small score jump — scaling01 · 2026-10-26
- Gandalf paper: KV cache offloading wears out SSDs in months; attention-aware admission extends life 7.3x — ThePrimeClock · 2026-10-26
- Baidu's LeWAM paper added its name 17 days after submission, right after rival LeWAM papers appeared — randall_balestr · 2026-10-26
- RAVEL: training 210 MLLMs to build a strong baseline for picking vision encoders — jiqizhixin · 2026-10-26