Skip to content

Wire Briefings Search

Fetching from the wire…

01

02

03

04

05

06

07

08

Wire Briefings Search

Entity trail

Puro-2B

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

0

Findings

1

Edges

3

Sources

1

Corpus findings

2026-08-31 / sources-researcherPuro-2B pretrains a 2B model on consumer RTX 5090s for under $6.9KPosted 2026-08-27, this report is an open pretraining recipe aimed squarely at the cost barrier: training Llama-3.2-3B runs over $1.5M and reproducing SmolLM3-3B needs over $700K. Puro-2B trains from scratch on up to 1.4 trillion tokens in FP8 on consumer-grade RTX 5090 GPUs, with the best model costing under $6.9K in compute and approaching Qwen2.5-1.5B under the authors' evaluation protocol. The savings come from stacked choices rather than one trick: hardware selection, low-precision training, hyperball optimization and curriculum model averaging. This is the recipe an independent builder can actually run.

Graph relationships

BENCHMARKED AGAINST
Qwen2.5-1.5B
Puro-2B approaches Qwen2.5-1.5B performance under the authors' evaluation protocol.
Source finding
BENCHMARKED AGAINST
SmolLM3-3B
Puro-2B achieves competitive cost compared to SmolLM3-3B pretraining.
Source finding
BENCHMARKED AGAINST
Llama-3.2-3B
Puro-2B pretraining approach costs less and approaches Llama-3.2-3B performance.
Source finding

Source trail

arXiv 2608.27370, via Hugging Face Daily Papers

Graph sources

entity graphfindings textkg edgeskg entities