Modelpublic

42ailab/MiroThinker-1.7-mini-GGUF

sourceHugging Faceapache-2.0updated 4d agoView on Hugging Face

0likes61downloads

Model Card

MiroThinker-1.7-mini · Fully Local Deep-Research Agent Model

![License: Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0) ![Runs Local](https://42model.com) ![Platform](https://42model.com) ![ModelScope](https://modelscope.cn/models/42ailab/MiroThinker-1.7-mini-GGUF)

English | 简体中文 · Also on ModelScope · 42ailab.com

Give it a question that needs repeated look-ups and several steps, and this model does the thinking and decides which tool to call next — entirely on your own computer. No cloud, free, private.

The model was built by [MiroMind AI](https://huggingface.co/miromind-ai/MiroThinker-1.7-mini) and released under Apache-2.0 (MiroThinker-1.7-mini, a Qwen3 mixture-of-experts architecture). This repo is not a new model — we converted that open model into a quantized build (GGUF) that runs offline on your computer, and verified that it loads and generates correctly on our own engine. The capability belongs to MiroMind AI; what we did was make it run locally.

1. The problem it solves

Some questions cannot be answered in one turn. You have to look something up, read it, then look up the next thing: checking where a claim came from, assembling scattered facts from a dozen pages into one table, following a lead down several levels. What this kind of work needs is not one polished answer — it is staying on track for a long time.

Doing that usually means handing your question, your search trail, and whatever internal material you read along the way to a cloud service. This model moves the half that thinks and decides back onto your machine:

—Sensitive topics — competitive research, due diligence, internal document work: you would rather not hand the trail to a cloud provider;
—Long chains — a single question that takes dozens of steps, where per-call cloud pricing adds up;
—Offline / intranet — restricted networks still need a local brain that can reason.

2. How it works

The capability comes from MiroMind AI's MiroThinker-1.7-mini. Its design points:

—Mixture of experts (MoE) — 30B parameters in total, but only a small fraction is used to produce each token, so memory use and speed are closer to a much smaller model;
—Very long context — up to a 256K-token window, enough to hold dozens of rounds of reading;
—Trained for multi-step tool use — upstream designed it for as many as 300 tool calls within a single task, not for answering in one shot.

What we did: converted the upstream weights into a single ~18.6 GB quantized file that runs offline across macOS / Windows / Linux, and kept the same Apache-2.0 license as upstream.

3. How well does it do

For benchmark results on the model itself, please refer to the upstream technical report (arXiv).

⚠️ One thing worth stating plainly: the deep-research benchmark numbers on the upstream model card are reported under the series name "MiroThinker-1.7", and in the same model table MiroThinker-1.7 is the 235B flagship while MiroThinker-1.7-mini is this 30B build — meaning those numbers belong to the flagship, and upstream does not publish separate numbers for this build. So we quote no benchmark figure at all here, and list only our own verification:

What we verified	Result
Size before / after quantization (original bf16 → this build)	61.1 GB → 18.6 GB, about 30% of the original
Does our engine recognize and load it	✓ architecture and chat template both detected correctly
Real-machine generation (Apple M3 Max)	✓ answers normally, thinking block closes properly, fluent Chinese

The upstream card lists English as the model's language, but its base is Qwen3, which is strong in Chinese — we verified fluent Chinese question answering in practice, so Chinese is listed here as well.

4. Limitations and what's next

—It needs a client that gives it tools. This model is at its best when it can look things up while it reasons, but search and web browsing are not inside the model — the client calling it has to provide them. Used as an ordinary chat model, it will feel less handy than a same-size chat model; that is not what it is for.
—Memory — a machine with 24 GB or more is recommended.
—Usable context length is bounded by your machine's memory; not every machine can open the full window upstream advertises.
—This repo currently ships only the Q4_K_M build (the balance point between size and quality).

5. How to download and use it

This build is packaged for 42model; getting it through 42model is the recommended path:

Desktop app In "Model Library" → "Research", download MiroThinker-1.7-mini.

Then start 42model's local service on your machine and point a tool-capable client (any agent client that speaks the OpenAI or Anthropic API) at it — this model will act as that client's local brain.

Files and license

File	Purpose
`MiroThinker-1.7-mini-Q4_K_M.gguf`	The model itself (4-bit quantized build)

The file's sha256 is shown on the repo's file page and can be verified yourself.

License: the model itself is MiroThinker-1.7-mini, © MiroMind AI, Apache-2.0 (official sources: Hugging Face · GitHub · ModelScope · technical report). This repo is a GGUF build converted from it and follows the same Apache-2.0; using it means agreeing to the upstream license terms.

Citation

For the model itself, please cite upstream MiroMind AI.

⚠️ The Citation section on the upstream model card still points at the 2025 v1.0 paper (arXiv:2511.11793), while this model belongs to the MiroThinker-1.7 series, whose technical report is arXiv:2603.15726 (MiroThinker-1.7 & H1: Towards Heavy-Duty Research Agents via Verification, MiroMind Team, 2026-03-16). Both identifiers appear on the upstream card; we reproduce the citation block upstream provides as-is and flag the inconsistency here so you can pick what you need — for work on the 1.7 series, use the latter.

bibtex

@article{miromind2025mirothinker,
  title={MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling},
  author={MiroMind Team and Bai, Song and Bing, Lidong and Chen, Carson and Chen, Guanzheng and Chen, Yuntao and Chen, Zhe and Chen, Ziyi and Dong, Xuan and others},
  journal={arXiv preprint arXiv:2511.11793},
  year={2025}
}

If this repo's GGUF build is useful to you, you may additionally cite:

bibtex

@misc{yang2026mirothinker17minigguf,
  title        = {MiroThinker-1.7-mini-GGUF: A Local Quantized Build of MiroMind AI's MiroThinker-1.7-mini},
  author       = {Yang, Zhiping},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/42ailab/MiroThinker-1.7-mini-GGUF}},
  organization = {42ailab},
  note         = {GGUF quantization and local-deployment packaging; the model itself is MiroMind AI's MiroThinker-1.7-mini (Apache-2.0). Contact: contact@42ailab.com}
}

About us

[42ailab](https://42ailab.com) — an AI research lab exploring the boundaries of intelligence. Grounded in cognitive science, we work toward a deep integration of AI and human intelligence, to truly understand and augment intelligence — carbon-based and silicon-based alike.

[42model](https://42model.com) — a high-performance local AI inference engine from 42ailab. It runs translation, transcription, recognition, conversation, coding and more on your own machine, free and private; and it can fine-tune a model of your own in the cloud and bring it back to run locally.