Samaya post-trains a 30B model to beat frontier AI on financial research at 11x lower cost
maithra_raghu · x · 2026-10-29
Samaya introduced Samaya Sterling, a model post-trained for end-to-end financial research that outperforms frontier models on its FrontierFinance benchmark, where the best frontier model (Claude Fable 5.1) only reaches 55.9% accuracy.
- The team post-trained Nemotron 3.5 Lightning (30B MoE, 3B active params) from 42.5% to 58.9% accuracy at $0.61 per run, vs. Fable 5.1's 55.9% at $6.89 — an 11x cost reduction.
- Training uses four stages: two supervised stages learning from eight teachers (financial tool use, research strategy selection, citable answer aggregation), followed by offline RL and online RL.
- A key finding is co-adaptation between model and harness: the harness supplies quality financial tools and compresses long traces so even a small model can sustain long-horizon research.
More from Models
- MiniMax M3.1 Flash scores 43 on Artificial Analysis index as skeptics predict flop — zacharynado · 2026-10-29
- LMArena launches an arena-style benchmark for agent capabilities — allenainie · 2026-10-29
- Gemini Carbon Arena benchmark results shared on Reddit, better than Argon — Bazinga8000 · 2026-10-29
- Two theories on why Google shipped Gemini 4 Carbon now — Angaisb_ · 2026-10-29
- Leaked checkpoints: Arena tester claims Gemini 4.0 Argon trails Sonnet ahead of next-week launches — zacharynado · 2026-10-29
- Don't extrapolate Google's progress: the Gemini speed meme hits RSI index — vitaliychiley · 2026-10-29