yousef-rafat/MaximusLLM
View on GitHub ↗ ★ 28 Python archived
High-throughput long-context LLMs. Scaling context via RandNLA and massive vocab capacity through MAXIS Loss and Fisher-SVD.
4 mentions on Hacker News by 1 person: 4 stories, 0 comments · first 2026-03-13 · latest 2026-03-16
20082026
-
▲ 2
Show HN: MaximusLLM – Train 262k-vocab LLMs on a single 16GB GPUstory by yousef_g · 2026-03-16 · 0 comments
-
▲ 1
Ghost Logits: Simulating missing partition mass in sampled softmax [pdf]story by yousef_g · 2026-03-15 · 0 comments
-
▲ 1
Show HN: MaximusLLM, Breaking transformer's O(N^2) and O(V) scaling bottlenecksstory by yousef_g · 2026-03-13 · 0 comments
-
▲ 1
MaximusLLM: High-Speed Architecture via Ghost Logits and Random Latent Attentionstory by yousef_g · 2026-03-13 · 0 comments