Benchmarks

← All benchmarks

Vals RSI (four tasks) — Post-training

Agents

Reference-normalized improvement of a provided Qwen3.6-35B-A3B checkpoint in 30 hours on sixteen RTX PRO 6000 GPUs, with internet, 27 public Finance Agent v2 examples and $100 external inference credits. One submitted checkpoint and generation setup is evaluated three times on 126 held-out tasks. Baseline 35.38%, reference 59.42%, theoretical best 99%. Component of the four-task RSI revision.

Top models (higher is better)