What it is
A 35B MoE agent model matches or beats some 1T models by scaling long-horizon trajectories and multi-domain agent skills instead of parameter count. Instead of scaling parameters, it scales agent horizon: builds a knowledge-action graph for 45K-token trajectories, then merges six domain specialists into one 35B student via domain-routed on-policy distillation with salient vocab alignment.
Key results
- SEAL-0: 56.4, above Kimi-K2.6 55.0 and DeepSeek-V4-pro 42.3
- IFBench: 80.6, above Kimi-K2.6 73.5 and DeepSeek-V4-pro 75.9
- FrontierScience-Olympiad: 79.0, above Kimi-K2.6 76.0 and DeepSeek-V4-pro 78.0
Long-horizon, tool-grounded training plus routed multi-teacher distillation can let a 35B agent reach 1T-class performance on agent benchmarks.