MiroThinker

MiroThinker is a series of open-source deep research agents developed by MiroMind AI, released under the Apache 2.0 license. The models are designed for long-horizon reasoning, web browsing, and prediction tasks, and are accessible both as downloadable model weights on HuggingFace and through the hosted MiroMind platform at dr.miromind.ai.

What It Is

MiroThinker is a family of tool-augmented language model agents built specifically for deep research and prediction. Unlike general-purpose LLMs, MiroThinker models are trained with an "interactive scaling" approach — systematically increasing the depth and frequency of agent-environment interactions as a third dimension of performance improvement, beyond model size and context length. The framework, called MiroFlow, is also open-source and supports flexible tool integration, trace collection, and benchmark evaluation.

Architecture and Scaling Approach

MiroThinker's core innovation is interactive scaling: training agents to handle deeper and more frequent tool interactions rather than simply scaling parameters or context. Key architectural properties include:

• 256K context window for long-horizon, multi-step analysis
• Up to 300–400 tool calls per task depending on version, enabling extended research trajectories
• Recency-based context retention (keep_tool_result) that preserves reasoning traces while focusing attention on recent observations
• DAG reasoning under MiroMind OS for stability in 100+ step tasks
• Available in 30B and 235B parameter scales (MiroThinker-1.7); earlier versions also offered 4B, 8B, 14B, and 32B variants

The agent framework integrates with external tools via MCP servers, including Google search (Serper), web scraping (Jina), and code execution (E2B sandbox).

Benchmark Performance

MiroThinker-1.7 (235B) achieves the following scores as reported by MiroMind:

• 74.0% on BrowseComp (English web browsing comprehension)
• 75.3% on BrowseComp-ZH (Chinese web browsing comprehension)
• 82.7% on GAIA-Val-165 (general AI assistant benchmark)
• 42.9% on HLE-Text (Humanity's Last Exam, text-only subset)

MiroThinker-1.7-mini (30B) achieves 72.3 on BrowseComp-ZH, which MiroMind claims sets a new SOTA among open-source models at that parameter scale. The proprietary MiroThinker-H1 agent achieves leading performance on BrowseComp and BrowseComp-ZH among both open-source and commercial models, per MiroMind's published results.

Update: MiroThinker-1.7 and H1

The most recent release, announced on March 11, 2026, introduces MiroThinker-1.7 (235B) and MiroThinker-1.7-mini (30B), alongside the proprietary MiroThinker-H1 agent. Key improvements in 1.7 include an enhanced post-training pipeline, more accurate stepwise reasoning, and support for up to 300 tool interactions per task. The accompanying technical report (arXiv:2603.15726) describes the verification-based approach for long-chain reasoning. Earlier milestones include MiroThinker-v1.5 (January 2026), v1.0 (November 2025), v0.2 (September 2025), and v0.1 (August 2025). The GitHub repository shows 8,306 stars and 641 forks as of the last recorded update.

Self-Hosting and Deployment

Researchers and developers can self-host MiroThinker using SGLang or vLLM for model serving, with the MiroFlow agent framework handling tool orchestration. The setup requires Python 3.10+, the uv package manager, and API keys for Serper (search), Jina (scraping), and E2B (code execution). Pre-configured YAML agent settings are provided for different task types and compute budgets. Quantized deployment options via llama.cpp and Ollama are also documented for CPU-optimized or lower-resource environments.

Hosted Platform

In addition to the open-source framework, MiroMind operates a hosted web application at dr.miromind.ai where users can run MiroThinker without self-hosting. The platform supports deep research report generation, document uploads (PDF, DOC, PPT, XLS, JPG), and a credit-based usage model. MiroMind also launched iOS and Android apps in March 2026, per a blog post on the company website.