What happened
In a November 10, 2026 announcement, Baseten describes work with Genspark to post-train Qwen3.5-397B into a specialized deep-research model. Genspark’s Deep Research agent plans a research strategy, runs searches, reads pages, and produces a cited report. Baseten says the resulting model is now the default model powering that agent in production. The company says it runs with NVFP4 weights on two B200 GPUs per replica.
On Genspark’s quality grader, the post-trained model scored 0.759, compared with 0.612 for the base model. Baseten also lists comparison scores: 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. Baseten says the model crawls five times more sources per user query than the base model and produces longer, better-cited reports. Those figures are results from Genspark’s evaluation rather than an independently published benchmark.
The teams began with Genspark’s existing evaluation infrastructure: representative research queries and an LLM grader served from Genspark’s endpoint. Baseten says they stress-tested the grader for blind spots and reward hacks, agreed on baselines and a held-out test set, and set conditions for production readiness. They also inspected agent trajectories rather than immediately training against a final-answer score. The announcement describes the work as two stages: research across the web, then synthesis of more than 50,000 tokens of collected evidence into a cited report.
To diagnose writing separately, the teams held research evidence fixed and compared a report from the base model with one written by a stronger writer model using the same evidence. On a small diagnostic panel, the score rose from 0.58 to 0.73. Baseten says this indicated that synthesis was a larger bottleneck than the base model’s research. Training then used different methods for the two stages: reinforcement learning for research behavior, and supervised fine-tuning on high-quality reports for writing, followed by reinforcement learning to integrate those behaviors.
The announcement also describes an iterative review process. Genspark compared candidate models with its production agent and a frontier model, examining coverage, depth, citations, trustworthiness, and performance on real production queries. The teams used manual review and a separate grader alongside the training grader. Baseten says reward shaping supplied more direct incentives for desired research behavior; for example, search targets were set per question rather than using one fixed target for easy lookups and complex analyses alike.
What the archive adds
New Runtime’s earlier coverage of fine-tuning GPT-OSS-20B for multilingual reasoning argued that fine-tuning should address a capability gap before optimizing an existing behavior. This is a new event on a related theme: Baseten and Genspark describe matching different training methods to research and synthesis behaviors.
Trend signal
This case does not yet fit a pattern New Runtime tracks. It does, however, reinforce the practical value of separating an agent’s evidence-gathering behavior from its report-writing behavior when building task-specific evaluations and training loops.
Why now
Baseten frames the engagement around a practical constraint: Genspark’s Deep Research workload is expensive at its volume when run on closed models, while the team wanted to train to its own quality bar. The announcement says Genspark was already using a fine-tuned open model, making the work an attempt to improve that existing approach rather than an initial move away from closed models.
Why it matters
The useful lesson for teams is methodological, not that one score settles which model is best. A research agent’s final report combines at least two capabilities: finding and selecting evidence, then writing from that evidence. Baseten’s account shows how a single end-to-end score can obscure where quality fails, and why separate diagnostics can guide different interventions.
It also highlights the risk of optimizing to a grader that ranks finished answers but offers weak direction during training. Genspark’s use of a held-out set, another grader, and manual checks matters because a model can improve on its training signal without improving the product experience. Teams considering a specialized open model should define task-representative evaluations and verify outputs outside the training loop before treating benchmark gains as operational gains.
Limits and uncertainty
The reported scores come from Genspark’s quality grader, and Baseten’s announcement does not establish that its test set or grading method is comparable to the evaluations behind the listed competitor scores. Baseten says the teams used a held-out set, another grader, and manual checks, but does not publish their sizes, score distributions, or independent replication results.
The five-times-more-sources figure is also reported by the partners; the announcement does not specify the query mix or the resulting changes in latency and cost. The production deployment is described by the partners and has not been independently audited here. The available comparison therefore supports a case-study claim about this implementation, not a general ranking of research models.
Skill to build now
For a research agent, make an evaluation that can score research and synthesis separately. Hold the collected evidence fixed when testing report writing, then compare research strategies on representative questions. Keep a held-out set and review samples manually so improvements to a grader score do not stand in for better reports.
Reader takeaway
Baseten and Genspark present a detailed account of specialized post-training reaching production, but the headline comparison remains a company-reported result. The reusable idea is to diagnose research and writing as separate skills, train them with different signals, and check gains with evaluations beyond the one used to shape training.