🤖 Small AI Matches Giants at Deep Research — Mind DeepResearch Proves It

What if a modest-sized AI could research as well as the biggest models on the planet?

Researchers at Li Auto just released Mind DeepResearch (MindDR) — a multi-agent system that uses ~30 billion parameter models yet rivals much larger systems on deep research tasks.

The secret? Three specialized AI agents working as a team:

• **Planning Agent** — decides what to search and in what order
• **DeepSearch Agent** — dives deep into the web for thorough information gathering
• **Report Agent** — synthesizes everything into a clean, readable report

Trained through a four-stage pipeline (supervised fine-tuning, search RL, report RL, and preference alignment), MindDR scores 42.8% on BrowseComp, 45.7% on BrowseComp-ZH, and 52.5 on DeepResearch Bench — competitive with far larger open-source agent systems.

What makes this exciting:

• **Efficiency** — comparable results at a fraction of the model size
• **Real-world deployment** — already running as a commercial product at Li Auto
• **Custom benchmark** — includes 500 real Chinese user queries for practical evaluation
• **Open methodology** — the four-stage training recipe is fully documented

Think of it as a detective squad: one plans the investigation, one gathers evidence, one writes the case report. Together, they outperform a single brilliant detective working alone.

This could be the turning point where powerful research AI becomes accessible to everyone — not just those who can afford the biggest models.