Curie: a from-scratch 17B model designed to run from SSD, 33 tokens/s on one CPU core
Just_Vugg_PolyMCP · reddit · 2026-09-12
A developer released colibrì, an engine treating SSD, RAM and GPU as a single memory hierarchy, plus Curie, a 17B model designed for it, asking why a model whose weights fit a $60 SSD should need a $30,000 machine.
Key facts
- colibrì can run models that don't fit in RAM or VRAM, including hundreds-of-billions-parameter models, on ordinary hardware.
- Most open models assume VRAM, so the author designed Curie from scratch: written in C, no inference framework, trained on a laptop without a GPU, weights living on SSD loaded by execution path.
- Demo reaches 33 tokens/s on a single CPU core, vs 3-5 tokens/s for dense 12B models on the same machine.
- Major caveat: Curie has seen only 11 million training tokens and isn't useful yet; billions of tokens are needed before proper evaluation. The alpha demonstrates the architecture; the model isn't public.
More from Infra
- SemiAnalysis: OpenAI reportedly negotiating ~1GW TPU commitment while seeking sites — dylan522p · 2026-10-28
- I tested 11 video upscaling methods for AI video: local, Topaz and cloud compared — lopas99 · 2026-10-28
- Thermo-backed CTO says AI is in its dial-up phase, touts disaggregated compute for an ultrafast 'broadband moment' — beffjezos · 2026-10-28
- DeepSeek-V4.1-Flash Q2 runs at 41 tok/s on one RTX 3090, 90% token match vs API — musicurgy · 2026-10-28
- USearch v3 RC ships after 1,000 days, its libraries now embedded in 550K-star projects — viglovikov · 2026-10-28
- SK Hynix says memory demand will outpace supply through 2030 — Chemical_Hawk_6307 · 2026-10-28