We’re releasing MiroMind Open eep Research (OR) v0.1, which we believe is the first full-stack, fully open-source deep research project—not just an agent, but also the model, dataset, and training/RL infra are open and reproducible. The agent framework (MiroFlow) reproduces 82.4 on GAIA validation; the model series (MiroThinker) reaches 60.2% on GAIA-Text-103. Looking for contributors + repro logs.

Why this matters

• Full-stack openness: most deep-research releases stop at the agent; OR opens all four layers: Agent (MiroFlow), Model (MiroThinker), ata (MiroVerse), Training/RL (MiroTrain / MiroRL).
• Reproducible numbers: • MiroFlow: GAIA validation maj. vote 82.4, pass@1 avg@3 72.2 (with setup details & scripts). • MiroThinker v0.1: 60.2% on GAIA-Text-103 (with both SFT & PO variants across 8B/14B/32B).
• Open data at scale: MiroVerse v0.1—147k+ full rollout trajectories (~1.9B tokens, 602k+ tool calls), built for tool-use/web-browsing agents.

What’s included

• MiroFlow (Agent framework) – multi-tool, sub-agent orchestration, MCP integration, benchmarking UI; detailed GAIA runs & scripts.
• MiroThinker (Model series) – agentic LLMs optimized for deep research; SFT/PO at 8B/14B/32B with evaluation guides.
• MiroVerse (ataset) – 147k+ verified trajectories across multi-hop QA, browsing, scientific reasoning; hybrid licensing noted on card.
• MiroTrain / MiroRL (Training & RL) – end-to-end post-training + MCP-first RL for tool-using agents.

Quick start (agent eval)

1. MiroFlow: clone, set keys (OpenRouter/Anthropic/OpenAI/Gemini, Serper, Jina, E2B), optional E2B ocker sandbox for stable repro; run GAIA scripts.
2. MiroThinker: pull model from HF or self-host via SGLang; run GAIA-Validation / GAIA-Text-103 / HLE / WebWalkerQA scripts.

Links

• Overview blog (tables & results): miromind.ai/blog/miromind-open-deep-research MiroMind
• Agent: GitHub.com/MiroMindAI/MiroFlow GitHub
• Models: GitHub.com/MiroMindAI/MiroThinker & HF collection GitHubHugging Face
• ataset: HF — miromind-ai/MiroVerse-v0.1 Hugging Face
• Training/RL: GitHub.com/MiroMindAI/MiroTrain & /MiroRL GitHub+1