Puro-2B: an open recipe trains a Qwen2-1.5B-beating LLM on RTX 5090s for just $4.4K
IgorCarron · x · 2026-09-25
Puro-2B announces a fully open pretraining recipe for a 2B LLM that can be trained from scratch on consumer RTX 5090 GPUs:
- $4.4K: a model that beats Qwen2-1.5B
- $6.9K: approaches Qwen2.5-1.5B level
- Data, pipeline, and hyperparameters are all public and reproducible
Igor Carron ties it back to his 2021 essay "The $1,000 GPT-3": just as genome sequencing costs collapsed far faster than Moore's law once competition intensified, small-model pretraining costs are now undergoing a similar commoditization — progress doesn't come from one technology's steady march, but from a diversity of challengers.
More from Infra
- Google to launch TPUs into space next week on Falcon 9 to test orbital AI data centers — McDonaghMatthew · 2026-09-25
- NVIDIA now tops the list of America's biggest businesses after a decade-long climb — lemire · 2026-09-25
- LithosAI uses GPU virtualization to push the Pareto frontier of agentic inference — JiaZhihao · 2026-09-25
- Analog chip runs LLM attention 100x faster than H100 using 70,000x less power, Nature paper claims — anselm · 2026-09-25
- Healthcare AI's GPU dilemma: balancing latency-sensitive clinical inference against batch research workloads — Arindam_1729 · 2026-09-25
- LexiPanel: Open-Source Control Panel Runs Local LLM, Image and Audio AI on Your Own GPU — W61k3r · 2026-09-25