HybridResearcher: a 4B deep research agent beats a 397B model

What the research found

HybridResearcher is a 4B agent trained on Qwen3.5-4B with synthetic cross-source trajectories and SFT plus RL, gaining 30.0 pass@8 points on HybridDeepResearch and beating the 397B Qwen3.5 model by 8.3 points, with no catastrophic forgetting on BrowseComp benchmarks.

Claims and how well the evidence supports them

1. HybridResearcher is a 4B agent built on Qwen3.5-4B 84%
2. HybridResearcher improves pass@8 by 30.0 points over its base model on HybridDeepResearch 98%
3. HybridResearcher surpasses Qwen3.5-397B-A17B by 8.3 points on HybridDeepResearch 98%
4. Training uses synthetic cross-source trajectories, SFT then RL 98%

Evidence

• huggingface.co https://huggingface.co/papers/2610.24899
• arxiv.org https://arxiv.org/abs/2610.24899

Actionability rated low.