Post
fry69
fry69.dev
did:web:fry69.dev
FYI: "GLM-5.3-Flash Architecture Notes" by @rasbt.bsky.social
- linear hybrid attention w/ 34 KDA + 11 MLA/DSA layers
- sparse 320B-A18B MoE backbone (~1/2 GLM-5.2)
- DeepSeek V4-style mHC residual path w/ 4 parallel streams
- native vision encoder (n/a)
-> https://sebastianraschka.com/blog/2026/glm-5-3-flash-architecture-notes.html
2026-08-28T06:31:47.771Z