Baseten and Genspark train open model to beat Opus 5.5

A higher score on Genspark's grader

The post-trained model scored 0.759 on Genspark's quality grader, compared with 0.612 for the base model. Baseten said that result was higher than 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.

According to Baseten, the model also crawls 5× more sources than the base model for each user query. It produces longer, better-cited reports and runs with NVFP4 weights on 2×B200 GPUs per replica.

Genspark's Deep Research agent plans a research strategy, runs searches, reads relevant pages, and returns a long, cited report. Genspark said the system handles long contexts and dozens of tool calls per question, while closed-weight models were expensive at its volume and could be changed only through prompting.

Training research and writing as separate tasks

Baseten and Genspark first examined the agent's trajectories instead of immediately starting reinforcement learning. They divided the work into a research phase, where the agent searches and reads sources, and a synthesis phase, where it turns 50K+ tokens of evidence into a cited report.

In one diagnostic comparison, the score rose from 0.58 to 0.73 when the base model's collected evidence was given to a stronger writer model. Baseten said this showed that report writing was the larger bottleneck, although both research and synthesis needed improvement.

The teams then used different training methods for the two phases. Reinforcement learning trained research behavior, using rewards to encourage strategies that produced better reports. For writing, supervised fine-tuning supplied high-quality reports written from the model's own research evidence. Baseten said a writing-focused reinforcement learning attempt initially stalled because nearly every report received a similar score. A supervised fine-tuning warm-up introduced behaviors including style, conciseness, and structured prose, after which reinforcement learning was used to integrate those behaviors with the model's research capabilities.

Production deployment and control over the model

Genspark evaluated candidate models through its own research queries, LLM grader, held-out test set, side-by-side comparisons, and manual reviews. The company assessed coverage, depth, citations, and trustworthiness, while feedback also addressed citation quality, corroboration across sources, and compliance with multipart instructions.

Baseten said Genspark's reviews led the teams to seek broader research and longer, more complete reports. The company also describes its Baseten Loops SDK as infrastructure for supervised fine-tuning and reinforcement learning, with visibility and control over the training process. Baseten Applied Research worked directly with Genspark on the training runs, diagnostic experiments, reward design, and debugging.

Kay Zhu, Genspark's co-founder, said the company would keep “the weights, the evaluations and the training code.” That gives Genspark control over a model trained on its own research tasks and quality standards, while the deployment demonstrates the company's stated goal of serving Deep Research at scale with an open model.