yousef-rafat/MaximusLLM
High-throughput long-context LLMs. Scaling context via RandNLA and massive vocab capacity through MAXIS Loss and Fisher-SVD.
2 best points · 0 comments · 4 Hacker News posts · 28 GitHub stars · View on GitHub
Hacker News discussions
- Show HN: MaximusLLM – Train 262k-vocab LLMs on a single 16GB GPU (2 points, 0 comments, by yousef_g, 2026-03-16)
- MaximusLLM: High-Speed Architecture via Ghost Logits and Random Latent Attention (1 points, 0 comments, by yousef_g, 2026-03-13)
- Show HN: MaximusLLM, Breaking transformer's O(N^2) and O(V) scaling bottlenecks (1 points, 0 comments, by yousef_g, 2026-03-13)
- Ghost Logits: Simulating missing partition mass in sampled softmax [pdf] (1 points, 0 comments, by yousef_g, 2026-03-15)