Model trained via Tinker beats rival on finance benchmark at 9% of the cost
simonguozirui · x · 2026-10-29
Samaya trained Nemotron 3.5 Lightning to beat Fable 5.1 on its finance research benchmark at just 9% of the cost, using the Tinker API. The recipe: first imitate 8 teachers, then run RL on its own mistakes — training the model together with the harness it operates in. The sharer notes how remarkably mature RL recipes have become over the past year and praises Tinker's API design for experimenting with new ideas.
More from Models
- Clover-35B-A3B-Preview, a diffusion MoE LLM, trends on Hugging Face — Diffuspace · 2026-10-29
- Quantized Ammu 1.0 guardrail: 17M-param model beats GPT-6-luna off-topic F1 by 10 points — kalyan_kpl · 2026-10-29
- Google AI gets cagey on who won the 2020 election — but answers 2024 and 2016 fine — JacquesThibs · 2026-10-29
- Researcher: Codex $500 plan burns credits faster than Anthropic's $20 sub — sethlazar · 2026-10-29
- Uncensored GGUF Quant of LiquidAI's LFM3-1.7B Trends on Hugging Face — AtomicChat · 2026-10-29
- Claude Pro vs ChatGPT Plus at $20: robotics dev crowdsources which holds up — Wild_Shower6854 · 2026-10-29