Instructions to use FirstPrinciplesLabs/quark with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use FirstPrinciplesLabs/quark with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="FirstPrinciplesLabs/quark")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("FirstPrinciplesLabs/quark")
model = AutoModelForMultimodalLM.from_pretrained("FirstPrinciplesLabs/quark", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use FirstPrinciplesLabs/quark with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "FirstPrinciplesLabs/quark"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FirstPrinciplesLabs/quark",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/FirstPrinciplesLabs/quark

SGLang

How to use FirstPrinciplesLabs/quark with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "FirstPrinciplesLabs/quark" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FirstPrinciplesLabs/quark",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "FirstPrinciplesLabs/quark" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FirstPrinciplesLabs/quark",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use FirstPrinciplesLabs/quark with Docker Model Runner:
```
docker model run hf.co/FirstPrinciplesLabs/quark
```

chengfp-eng commited on 14 days ago

Commit

a524ab1

verified ·

1 Parent(s): 97da9ad

Model card: general description of the training data

Browse files

Files changed (1) hide show

README.md +12 -23

README.md CHANGED Viewed

@@ -17,11 +17,11 @@ tags:
 # Theo-Quark
-Theo-Quark is a 9B reasoning model for physics, science and maths, built by [First Principles](https://firstprinciples.com/). It is
 [Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) with additional post-training: supervised fine-tuning on
-about 200,000 verified worked solutions, mostly physics.
-It outperforms the model it started from on every hard benchmark we ran at 131K maximum output token limit, and reaches
 graduate-level science scores comparable to much larger open models.
 ## Results
@@ -40,7 +40,8 @@ is graded by Qwen3.5-27B as judge.
 Both columns were measured by us with the same setup.
-**Compared with other open models** (scores as reported by each developer):
 | Model | Size | GPQA Diamond | AIME 2025 | HLE |
 |---|---|---|---|---|
@@ -54,28 +55,16 @@ Both columns were measured by us with the same setup.
 | DeepSeek-R1-0528 | 671B | 81.0 | 87.5 | 17.7 |
 | Kimi K2 Thinking | 1T | 84.5 | 94.5 | 23.9 |
-— = not reported. Sources: each model's official model card;
 ## Training data
-198,319 single-turn questions with worked solutions. The questions come from the datasets below; the
-answers were newly generated by the teacher model and kept only when correct.
-| Source | Content | Examples |
-|---|---|---|
-| First Principles physics set (from Nemotron science) | physics | 93,164 |
-| [nvidia/Nemotron-SFT-Science-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Science-v2) | other sciences | 28,965 |
-| [nvidia/Nemotron-SFT-Math-v4](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Math-v4) | maths | 29,776 |
-| [nvidia/AceReason-1.1-SFT](https://huggingface.co/datasets/nvidia/AceReason-1.1-SFT) | maths | 28,892 |
-| [nvidia/AceReason-1.1-SFT](https://huggingface.co/datasets/nvidia/AceReason-1.1-SFT) | code | 17,522 |
-- **Teacher model:** all answers were generated by
-  [Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next).
-- **Correctness:** maths and science answers were checked against reference answers (option letter, exact
-  match, math-verify, numeric checks, then an LLM judge) and only correct ones kept. Code was not checked.
-- **Modes:** 80% of examples include the full reasoning (thinking mode), 20% answer directly (non-thinking).
-- **Decontamination:** questions sharing long word sequences (13-grams) with GPQA, AIME 2025, MATH-500,
-  MMLU-Pro, SuperGPQA or HLE were removed.
 ## Training

 # Theo-Quark
+Theo-Quark is a 9B reasoning model for physics, science and maths, built by First Principles. It is
 [Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) with additional post-training: supervised fine-tuning on
+our curated physics, maths and coding mix.
+With room to think, it outperforms the model it started from on every hard benchmark we ran, and reaches
 graduate-level science scores comparable to much larger open models.
 ## Results
 Both columns were measured by us with the same setup.
+**Compared with other open models** (scores as reported by each developer; they use their own settings,
+mostly with shorter output limits, so small differences are not meaningful):
 | Model | Size | GPQA Diamond | AIME 2025 | HLE |
 |---|---|---|---|---|
 | DeepSeek-R1-0528 | 671B | 81.0 | 87.5 | 17.7 |
 | Kimi K2 Thinking | 1T | 84.5 | 94.5 | 23.9 |
+— = not reported. Sources: each model's official model card; Qwen3-8B's scores are from NVIDIA's
+Nemotron Nano 9B v2 card.
 ## Training data
+Theo-Quark was fine-tuned on our curated mix of about 200,000 physics, science, maths and coding problems,
+each with a worked solution written by [Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next).
+Solutions were checked for correctness, and problems overlapping the benchmarks above were removed. Most
+examples include the full reasoning; a fifth are direct answers, so the model keeps both thinking and
+non-thinking modes.
 ## Training