Home Open Source Article

MiroThinker v1.5 Open-Source Agent Achieves 80.8% on GAIA Benchmark

TL;DR

MiroThinker v1.5 introduces interactive scaling for agentic AI, handling 400+ tool calls with 256K context windows, surpassing commercial agents on multiple benchmarks.

Key Points

  • Achieves 80.8% on GAIA-Val-165, 71.5% on BrowseComp-ZH, outperforming Kimi-K2-Thinking with 1/30th the parameters
  • Introduces interactive scaling as third performance dimension beyond model size and context length
  • Supports 256K context window and up to 400 tool calls per task in 30B and 235B parameter scales
  • Includes MiroFlow agent framework (82.4% GAIA score) and MiroVerse training dataset (147k samples) on HuggingFace

Why It Matters

Open-source agentic AI is rapidly closing the gap with commercial systems. MiroThinker demonstrates that careful architectural choices around tool-use scaling can achieve competitive performance at lower computational cost, making advanced research agents accessible to developers without massive budgets. The interactive scaling approach offers a new optimization dimension for the community to explore.
View on GitHub

Source: github.com