Found first: a primary source the press has not covered yet.
Researchers at Tsinghua University have trained a 2-billion-parameter language model from scratch on consumer hardware for less than $6,900 and published the complete training recipe under Apache 2.0. The paper, submitted to arXiv on August 27, 2026, releases weights, data, code, and a fitted cost scaling law they call the Puro Cost Scaling Law.
What the source says
The Tsinghua team trained Puro-2B on up to 1.4 trillion tokens using FP8 precision on RTX 5090 GPUs. Their best model cost less than $6,900 to train. The fitted Puro Cost Scaling Law projects that about $4,400 is sufficient to reach Qwen2-1.5B performance, which the paper notes is less than $5,090. Puro-2B performance approaches Qwen2.5-1.5B under their evaluation protocol. For comparison, the paper estimates reproducing SmolLM3-3B would cost over $700,000.
Why it matters
Full LLM pretraining has been functionally closed to most academic labs, limited to industry teams with large compute budgets. Puro-2B is the first complete, reproducible open recipe for training a competitive small LLM on consumer hardware, with a projected cost below the retail price of the GPU used to run it. The paper also runs controlled studies on how pretraining data curricula shape downstream performance after post-training, adding methodological substance to the cost result. The Apache 2.0 license means any researcher can build directly on the weights and recipe without restriction.