Genspark's post-trained Qwen beats Opus on research
Genspark and Baseten post-trained the open Qwen3.5-397B into a Deep Research model that edges past closed frontier models on Genspark's own grader, and it now serves production traffic on two B200 GPUs per replica.
Baseten described how its Applied Research team worked with AI workspace company Genspark to post-train Qwen3.5-397B for Genspark's Deep Research agent. On Genspark's internal quality grader, the model scores 0.759, up from 0.612 for the base model. Baseten reports that this is ahead of Sonnet 5.5 (0.742), Opus 5.5 (0.738), Kimi K3 (0.724), DeepSeek V4 Pro (0.675) and GPT-5.6 Sol (0.637).
The scores come from Genspark's own LLM grader and a held-out test set, not from a public benchmark. Baseten says it stress-tested the grader for blind spots and reward hacks before training against it.
According to the post, the tuned model crawls five times as many sources per query as the base model and produces longer reports with better citations. It runs with NVFP4 weights on 2×B200 GPUs per replica and already powers Genspark's Deep Research agent in production. Before this work, Genspark was running the agent on an open model it had fine-tuned itself. The company's stated reasons for using open weights were cost at its volume and the ability to change the weights rather than only the prompts.
On method, the team treated research and synthesis as separate problems. In a diagnostic test, a stronger model rewrote reports from the same collected evidence, and scores rose from 0.58 to 0.73. That pointed to report writing as the main bottleneck. The team trained the research phase with RL. Its first attempt to train writing with RL stalled because the reports scored too uniformly to give a useful signal, so the team added a supervised fine-tuning warm-up before further RL.
Training ran inside Genspark's own agent harness and tools. Baseten did not disclose training compute, serving costs or latency figures.