Quark

Quark

A 9B reasoning model for physics, science and maths, fine-tuned from Qwen3.5-9B by FirstPrinciples. For researchers who want step-by-step derivations and a second opinion, locally or offline.

Made by FirstPrinciplesLabs 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: Quark is a 9B reasoning model from FirstPrinciples for physics, science and maths. It starts from Qwen3.5-9B and adds supervised fine-tuning on a curated mix of physics, maths and coding data. The maker also publishes an 8-bit Q8_0 version for llama.cpp.

Who it suits: Researchers and advanced students who want step-by-step reasoning on physics, science and maths questions.

Industry | Who uses it | How they use it | Example
Physics research | Physicists | Work through derivations and estimates | A step-by-step derivation of a result
Mathematics | Mathematicians | Check calculations and proofs | A second opinion on a worked solution
Science labs | Research teams | Keep sensitive work on their own machine | Offline analysis of unpublished data
Education | Advanced students | Study worked reasoning on hard problems | A graduate-level physics question

You can skip it if you need a general chat model. For results that matter, the card says to verify the answer independently.

Where to use it:

• Hugging Face Transformers, through the Safetensors weights
• llama.cpp, through the Q8_0 version, which the card says runs on a 16 GB Mac
• Theo Platform, a hosted research platform the maker runs
• Windows, macOS or Linux, with Python for the download step

Use case | Who | What they get back
Solving problems | Researchers | Derivations, estimates and calculations with reasoning
Checking work | Researchers | A second opinion on a result
Private research | Labs | Answers produced offline on your machine

How to set it up:

• Python and pip on your computer
• About 19.3 GB of space for the BF16 weights, or about 4.8 GB at 4-bit by estimate
• Transformers, or llama.cpp for the Q8_0 version

1. Install the Hugging Face hub package.
2. Download the model files with the hf command.
3. Follow the usage section of the model card to run it.

pip install -U huggingface_hub
hf download FirstPrinciplesLabs/quark

How to use it: Load the model in Transformers, or the Q8_0 file in llama.cpp, and ask a physics, science or maths question. The card says its benchmarks ran in thinking mode, with output up to 131,072 tokens.

hf download FirstPrinciplesLabs/quark

Example: You paste a physics derivation and ask the model to check each step. It walks through the reasoning and flags a step that does not follow, and you verify any result that matters separately.

For newcomers and experts: If you are new, start with a problem whose answer you already know and read the reasoning it writes. Treat its output as a draft to check.

If you know the field, the card reports GPQA Diamond of 87.4 for the bf16 weights, against 81.6 for Qwen3.5-9B. It says Q8_0 matches bf16 on every benchmark listed. It also says Quark sits on the size-versus-score frontier for text-only Humanity's Last Exam.

Qwen3.5-9B base model
Supervised fine-tuning on physics, maths and code
Quark in bf16
Q8_0 version for llama.cpp

Title: GPQA Diamond accuracy in percent
Qwen3.5-9B: 81.6%
Quark bf16: 87.4%
Quark Q8_0: 87.8%

Good to know:

• License: Apache 2.0
• Cost to run: your own hardware, while Theo Platform needs a sign-up whose price the card does not give
• Data it reaches: the files you download, and your questions if you use the hosted platform
• Not for: results that matter without independent checking
• Language: English only on the record

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 9.7B
• Weights: about 19.3 GB in BF16; about 4.8 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 107
• Likes: 7
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.5-9B
• Quantized versions: 1
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, physics, science, reasoning, math, qwen3.5, conversational

Links

• Model: https://huggingface.co/FirstPrinciplesLabs/quark
• Files: https://huggingface.co/FirstPrinciplesLabs/quark/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download FirstPrinciplesLabs/quark

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