lintool's Gaggle Matches SFT'd Qwen With Just a Handful of GPUs on nanochat Backbone
lintool · x · 2026-10-20
Jimmy Lin (lintool) shares two components of his gaggle project: a nanochat pretrained backbone on Hugging Face (built on karpathy's climbmix dataset, MIT-licensed, runnable via vLLM/ollama/llama.cpp) and a reranker. He claims performance on par with SFT'ing Qwen or Gemma, just a bit below listwise GPT-6.1 — all achieved with only a handful of GPUs.
More from Models
- REAX inference engine runs Perplexity Decider up to 6x faster, free for individuals — airesearch12 · 2026-10-20
- SKY-001 debuts as first 'Engineering World Model,' claiming 125x speedup over GPT-6 on CAD — vasuman · 2026-10-20
- Smart glasses test: Octen search at 60ms, model generation is the bottleneck; Mercury 3 wins — Scobleizer · 2026-10-20
- BitNet CPU inference is silently wrong: perplexity 6x reference, one-line fix works — matheus-c-feitosa · 2026-10-20
- Space Bunny stealth model revealed as MiniMax M3.1 Flash: 428B-A23B, weights Thursday — BTA_Labs · 2026-10-20
- Clairic is live on Workers AI; open-weights release planned once the architecture matures — michellechen · 2026-10-20