TinyFish runs its web-native agent model on Databricks AI Runtime
TinyFish is developing a specialized web-native model for enterprise agents that work across the live web, and has consolidated its post-training, evaluation and production serving on Databricks: managed GPUs via AI Runtime, evaluation with MLflow, and deployment through Model Serving. The team started with supervised fine-tuning and is moving toward Direct Preference Optimization. On its internal go/no-go benchmark, the model completed 1,691 of 2,178 web-agent tasks for a 77.6% task-success rate, with a separate LLM judge reviewing the request, action history and final result; partial completion or incorrect answers count as failures. Over a recent 15-day period, TinyFish recorded 294,781 agent runs involving the model across 6.4 million tool steps, at an estimated serving cost of $0.088 per run and $0.0040 per tool step — covering only that model, not the full agent stack. TinyFish already serves its search reranker and query rewriter on Databricks and uses Unity Gateway to centralize access to external foundation models.
It illustrates a pattern where agent teams converge post-training, evaluation and deployment onto one platform, and judge models by whole-task success rather than single-response quality.