MiroThinker: Open-Source Deep Research Agent by MiroMind AI

MiroThinker is an open-source deep research agent developed by MiroMind AI, focused on tool-augmented reasoning, multi-step long-horizon reasoning, and fact verification — not just another chat model.

Rather than a plain LLM, it works like a human researcher: propose hypotheses → retrieve evidence → iteratively verify → self-correct → deliver credible conclusions. It is especially suited to knowledge-intensive tasks such as complex research, forecasting, competitive analysis, and technical due diligence.

Core Innovation: Interactive Scaling (a "Third Dimension")

Traditional AI scaling relies on model size (parameters) + context length. MiroThinker adds a third dimension, Interactive Scaling:

• Trained with reinforcement learning (RL) to actively handle deeper, more frequent agent-environment interactions.
• Supports a 256K context window and up to 400-600 tool calls per task (far beyond previous open-source agents).
• Sustains multi-turn reasoning, error correction, and multi-source evidence synthesis, avoiding the "the more it thinks, the worse it gets" failure mode of classic test-time scaling.

The latest release (March 16, 2026: MiroThinker-1.7 & MiroThinker-H1) upgrades to a Verification-Centric architecture:

• Dual-layer verification system (local + global): real-time evaluation of every step during reasoning, plus whole-trajectory audits to ensure the evidence chain is complete.
• Significantly improved reliability, especially on long-chain, multi-step research tasks.

Versions & Models

• v1.5 (late 2025): 30B / 235B parameters (MoE architecture, fewer active parameters) — small models already beating 1T-class giants.
• v1.7 / 1.7-mini: more efficient; 1.7-mini (only 3B active parameters) reportedly surpasses GPT-5 on multiple benchmarks.
• H1: flagship heavy-duty research agent with the full verification system.

Benchmark highlights:

• BrowseComp / BrowseComp-ZH: leading among open-source, near or above GPT-5.
• GAIA, HLE, FinSearchComp and others: SOTA or highly competitive, with especially impressive small-model results.

How to Use

1. Online (easiest): Visit https://dr.miromind.ai/ to chat with MiroThinker directly (Pro tier supported). A mobile app (MiroMind App) is also available.

2. Local deployment / self-hosting (for open-source enthusiasts):

• Model weights: Hugging Face (under the miromind-ai organization)
• Code & framework: GitHub (MiroMindAI/MiroThinker)
• Runs locally with Ollama / LM Studio — fully private, controllable, no black box.
• Companion ecosystem: MiroFlow (workflow orchestration), MiroVerse (training datasets), MiroTrain (training toolchain).

3. License: MIT — code, weights, and datasets are all open.

Real-World Feedback (from recent discussions)

• Users report strong performance on client negotiation scripts, competitor research, and deep investigation, clearly outperforming ChatGPT / Claude / Gemini / Perplexity in scenarios requiring "think-verify-iterate" loops.
• Drawbacks: the online version's quota is relatively expensive (limited daily free usage, high per-task consumption on Pro); app stability still needs improvement.
• Small models also run very well locally — excellent value for money.

TL;DR: MiroThinker represents a new direction for open-source agents — not just stacking parameters, but thinking, verifying, and interacting — bringing reliable deep research into the hands of everyday developers. The latest 1.7 / H1 series pushes this idea to a new level. Try it at dr.miromind.ai or pull the weights and run it locally.