Hugging Face weight dump shows a math SFT with AIME 2025 gains at 16k context
kastnerkyle · x · 2026-08-04
A Hugging Face post shares the weights for a light math SFT cold start used for RL runs, trained on 500M tokens from V4 Pro math reasoning and trajectories 6k–12k tokens long.
It also reports AIME 2025 results for the model:
- At 8k context: avg@16 = 16.04%, pass@16 = 43.33%
- At 16k context: avg@16 = 17.71%, pass@16 = 46.67%
The author says this was the first training run that didn’t come with the usual infrastructure trauma such as poor throughput, kernel failures, or loss spikes.
More from Models
- Anthropic says Fable 5 reproduced 5 of OpenAI’s 10 Astra math advances in 24 hours — EricBuess · 2026-08-04
- OpenAI releases Lean certificates and walkthroughs for its new math results — OpenAI · 2026-08-04
- OpenAI says advances in foundational math could ripple into GPS, weather and medicine — OpenAI · 2026-08-04
- OpenAI says a next-gen model produced 10 new results on open math problems for about $2,000 — OpenAI · 2026-08-04
- Google may launch Gemini 3.5 Pro this week amid fierce price pressure — haider1 · 2026-08-04
- Hugging Face adds hosted inference for Kimi K3 through Fireworks and Together — MaziyarPanahi · 2026-08-04