From the source
Baseten Applied Research worked with
Genspark to post-train Qwen3.5-397B into a specialized Deep Research model that now powers Genspark's Deep Research agent in production.
- On Genspark's own quality grader the post-trained model scores 0.759, up from 0.612 for the base model, and above Opus 5.5 at 0.738, Sonnet 5.5 at 0.742, GPT-5.6 Sol at 0.637, Kimi K3 at 0.724 and
DeepSeek V4 Pro at 0.675.
- Genspark says it moved to an open model because frontier closed-weight models are expensive at its volume and allow only prompt changes, not weight changes.
- Training ran against Genspark's own agent harness and tools, so the model learned in the environment it would serve in.
- Diagnosing the agent's trajectories, the teams found synthesis was the larger bottleneck: holding research fixed and swapping in a stronger writer model raised the score from 0.58 to 0.73 on a small diagnostic panel.
- The model crawls 5× more sources per user query than the base model and writes longer, better-cited reports, and runs with NVFP4 weights on 2×B200 GPUs per replica.
- Genspark keeps the weights, evaluations and training code, according to co-founder Kay Zhu.





