There’s an inherent limit in training tool-using web agents: either you clone trajectories from a fixed teacher model, or you rely on sparse rewards from reinforcement learning, with little supervision for long interactions. DeepSearch-Evolve sidesteps the problem with self-distillation inside a verifiable environment, producing a competitive 9-billion-parameter agent that never asked for help from larger models.

At its core is DeepSearch-World, a deterministic and reproducible environment featuring 420,000 multi-hop QA tasks built from entity-level random walks. Every interaction leaves verifiable traces: the system can check whether the answer is correct, whether pages were actually read, whether the reasoning holds up. It’s not just a benchmark—it’s a playground designed for self-improvement, where the agent generates trajectories, filters them, mixes data, and undergoes iterative fine-tuning. That’s where DeepSearch-Evolve shows its edge: without distillation from more capable models, the DeepSearch-World-9B agent reaches 31.2% on BrowseComp, 61.5% on GAIA, and 93.4% on HotpotQA, matching leading open-source agents.

The true significance goes beyond the numbers. Organizations handling sensitive data—banks, healthcare, public administration—know that sending a prompt to an external LLM means exposing critical information. DeepSearch’s approach points in a different direction: a continuous improvement loop running entirely on your own infrastructure, in an environment where every step is auditable. There’s no need for a giant teacher model, nor to forward queries to third-party servers. Self-distillation also reduces dependency on model providers, a key factor when evaluating total cost of ownership and data sovereignty.

A 9-billion-parameter model isn’t nano-scale: real-time inference requires GPUs with sufficient VRAM, and the full iterative training loop demands careful hardware planning. Yet it’s a size that makes on-premise deployment feasible with a reasonable investment, far from the clusters needed for hundred-billion-parameter behemoths. For organizations already weighing self-hosted stacks, this framework shows that autonomous evolution is no pipe dream: you can build a search agent that learns from its own mistakes, without ever stepping outside the corporate perimeter.

DeepSearch signals a structural shift in the web agent landscape: from reliance on a few centralized models toward a fragmented, personalized training ecosystem. The full release—code, dataset, model, and validation set—aims to accelerate research on self-improving agents. For those designing on-premise architectures, the message is clear: verifiability and self-training are concrete levers to hold performance, control, and regulatory compliance together.