MiroThinker-v1.0-8B

MiroThinker-v1.0-8B

An open research agent model that reasons over long contexts and calls tools many times per task to look up information. The 8B size, built on Qwen3-8B, suits developers who run research agents on their own hardware.

Made by miromind-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MiroThinker-v1.0-8B is the 8B size of MiroThinker v1.0, an open-source research agent model built on Qwen3-8B. It is trained to work with tools and the outside world, using feedback from each step to correct errors. The card calls this interactive scaling. It supports a 256K context window and up to 600 tool calls per task.

Who it suits: It suits developers and researchers who want a research agent they can run themselves, with deep multi-step searching and analysis, instead of relying on a hosted service.

Industry | Who uses it | How they use it | Example
Research | Analysts | Let an agent search, read and cross-check sources over many steps | Gathering evidence on a question from several pages
Software | Agent developers | Build a research agent on an open 8B model | Wiring the model to a search tool and a reader
Academia | AI researchers | Study how deeper tool use affects results | Comparing short and long interaction runs
Business | Market teams | Run long information-seeking tasks | Compiling background on a company

Skip it if you want a plain chat model with no tools, since this one is trained around tool calls and long research runs.

Where to use it:

• Agent setups that give a model search and reading tools
• Python scripts using the transformers library
• Windows, macOS and Linux, on a machine that can hold about 16.4 GB of F16 weights

Use case | Who | What they get back
Deep research question | Analysts | A multi-step investigation ending in an answer
Long document analysis | Researchers | Reasoning over up to 256K tokens
Agent building | Developers | A model that handles many tool calls in one task

How to set it up:

• Python, with the transformers library
• About 16.4 GB of disk space for the F16 weights, or about 4.1 GB at 4-bit as an estimate
• The MiroThinker tools and workflows the card refers to, plus any search tools you connect

1. Install the Hugging Face tools.
2. Download the model files to your computer.
3. Connect the model to an agent setup with tools, following the MiroThinker project notes.

pip install -U huggingface_hub
hf download miromind-ai/MiroThinker-v1.0-8B

How to use it: You give the agent a research question and let it call tools in a loop, reading what comes back and deciding the next step. The part of the card available here shows no run command, so use the project's own tools and workflows for the exact steps.

Tool | What it does
Tool calls | Let the model look things up, up to 600 times per task
256K context window | Holds long working notes and fetched pages
Agent workflow | Loops between reasoning and fresh information

Example: You ask the agent to investigate a question that needs many pages. It searches, reads, checks what it found against other sources, and keeps going until it can answer. The card says deeper runs like this are where the model gains most.

For newcomers and experts: Think of it as a model that researches rather than only answers from memory. A good first try is to pair it with one search tool and ask a question that needs two or three lookups.

If you know the field, the card says the model is trained to handle deeper and more frequent agent and environment interactions, a third scaling axis next to size and context. It reports that the reinforcement-learned variant runs far longer interaction chains than the supervised one. The family has 8B, 30B and 72B sizes.

Receive a research question
Reason and choose a tool
Call the tool and read the result
Fix errors using the feedback
Answer when the evidence holds up

Title: Scores the card reports for the v1.0 family
HLE-Text: 37.7%
BrowseComp: 47.1%
BrowseComp-ZH: 55.6%
GAIA-Text-103: 81.9%

Good to know:

• Scores: the card gives these results for MiroThinker v1.0 in general, not per size, so the 8B model may score lower.
• Evaluation: the card says access to Hugging Face was blocked in the tools during testing, to prevent answer leaks.
• Sizes: 8B, 30B and 72B, with the 8B built on Qwen3-8B.
• License: MIT.
• Cost to run: none named for the model, though any search tools you connect may have their own costs.
• Data it reaches: it downloads model files from Hugging Face, and in an agent setup it reaches whatever web or tools you connect.
• Not for: users who need a model without tool calls, since the design centers on them.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 8.2B
• Weights: about 16.4 GB in F16; about 4.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 509
• Likes: 69
• License: mit
• Languages: en
• Based on: Qwen/Qwen3-8B
• Quantized versions: 9
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3, text-generation, agent, open-source, miromind, deep-research, conversational

Links

• Model: https://huggingface.co/miromind-ai/MiroThinker-v1.0-8B
• Files: https://huggingface.co/miromind-ai/MiroThinker-v1.0-8B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download miromind-ai/MiroThinker-v1.0-8B

Then follow the usage in its model card. Nothing in this kit runs by itself.