Deltafin
ARGODRIVE Deltafin: Kimi K3 (2.8T MoE) streamed from SSDs on Apple Silicon — fork of gavamedia/deltafin with the ARGODRIVE storage work and benchmark package -…
About Deltafin
TL;DR — Kimi K3 (2.8T-parameter MoE, 1.45 TB of expert weights) running on one M5 Max MacBook Pro with 128 GB, experts streamed from four SSDs.
1.00 tok/s steady decode over a 512-token answer; 1.13 over 128; 0.96 on the public 17-token prompt (upstream reported 0.68).
The honest limit: a 512-token prompt takes ~6.3 minutes to its first token. Cause found (prefill re-reads each layer's experts 8×), fix planned, not built.
Useful findings: one drive gives ≈52% of four-drive speed, two ≈73%, three ≈90% — the slowest of each layer's 16 reads sets the pace, not total bandwidth.
Discussion
Sign in to join the discussion.
Loading comments…
Tagged
Alternatives to Deltafin
Tools in the same space, ranked by how they are performing in the directory.
Ebm Lens
Agentic RAG pipeline for biomedical evidence discovery: query planning and multi-round search across twelve public databases plus citation-graph expansion, ran…
/Research
Collections
Drag links, text and images from any page into a side panel and keep them in collections for a trip, a purchase, a project.
/Developer ToolsThe Edge That Wasn
A memoir of retail FX algorithmic trading: 81 experiments, a 55,000-pip lookahead bug, four retractions, and the answer was no.
Paid/Developer Tools- Open source/Developer Tools
Gizmodo
Two recent studies have suggested that GLP-1 drugs can reduce the occurrence or severity of certain infections.
/Developer Tools
FrameThrower
Search the world’s largest film-still library for movie screenshots and cinematography references, learn from lighting and composition, and build lookbooks for
Paid/Developer Tools
