MiroThinker: Open-Source Deep Research Agent from 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 — rather than being a plain chat model.

It is not just an LLM: it works like a human researcher — form 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: a "third dimension" of scaling (Interactive Scaling)

Traditional AI scaling relies only on model size (parameters) + context length. MiroThinker adds Interactive Scaling:

• Trained with reinforcement learning (RL) to handle deeper and 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).
• Can reason across many turns, correct mistakes, and integrate multi-source evidence, avoiding the "overthinking goes wrong" failure mode common in traditional test-time scaling.

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

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

Main versions and models

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

Benchmark highlights (partial):

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

How to use it

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

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

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

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

Hands-on feedback (from recent discussions)

• Many users report it performs extremely well on client negotiation scripting, competitor research, and deep investigations, clearly better than ChatGPT / Claude / Gemini / Perplexity — especially in scenarios requiring "think–verify–iterate."
• Drawbacks: the online version's quota is relatively expensive (limited free daily usage; Pro-tier consumption per query is high); app stability still needs improvement.
• The small models also run very well locally — excellent value.

One-line takeaway: MiroThinker represents a new direction for open-source agents — not stacking parameters, but being able to think, verify, and interact — bringing "reliable deep research" into the hands of ordinary developers. The latest 1.7 / H1 series pushes this idea to a new level; try it at dr.miromind.ai or pull it down and run it locally.