Distilled 120B Model Beats Kimi in Finance Reasoning at 1/60th Cost
ycombinator · x · 2026-07-31
Following Semafor's report questioning the nationality of an American model distilled from a Chinese one, Cyril Gorlla shared benchmark data. At the 8k token budgets typically used in production, their 120B model scores 83.61% on FinanceReasoning, outperforming Kimi K3 (81.93%) and Inkling (65.13%). Running on a single H100, it achieves this at 62 to 160x lower cost per query. However, at unlimited budgets, the big models still win on raw accuracy.
More from Models
- OpenAI's GPT-5.6 Self-Optimizes: Slashes Serving Costs by 20% — tszzl · 2026-07-31
- Bypassing Pangram v4 AI Detection: Short Poetic Verses Slip Through — ctjlewis · 2026-07-31
- Scholars Debate Scaling Laws: Are Models Less General Despite Growing Stronger? — davidmanheim · 2026-07-31
- Neutrino-8B Hits HF Trending with Sub-2-bit Ternary Quantization — FermionResearch · 2026-07-31
- Kwaipilot KAT-Coder-V2.5 Trends on Hugging Face for Agentic Coding — bartowski · 2026-07-31
- True Positive Weekly #171: The AI Economy, SynthID Watermark, and Kimi K3 Weights — burkov · 2026-07-31