Instructions to use feyninc/sqrl-4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use feyninc/sqrl-4b with Transformers:

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

pipe = pipeline("text-generation", model="feyninc/sqrl-4b")
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)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("feyninc/sqrl-4b")
model = AutoModelForMultimodalLM.from_pretrained("feyninc/sqrl-4b", 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=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use feyninc/sqrl-4b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "feyninc/sqrl-4b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "feyninc/sqrl-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/feyninc/sqrl-4b

SGLang

How to use feyninc/sqrl-4b 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 "feyninc/sqrl-4b" \
    --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": "feyninc/sqrl-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "feyninc/sqrl-4b" \
        --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": "feyninc/sqrl-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use feyninc/sqrl-4b with Docker Model Runner:
```
docker model run hf.co/feyninc/sqrl-4b
```

bhavnicksm commited on 22 days ago

Commit

c5d327c

verified ·

1 Parent(s): bcc6361

Upload README.md with huggingface_hub

Browse files

Files changed (1) hide show

README.md +17 -13

README.md CHANGED Viewed

@@ -34,26 +34,30 @@ It was trained in two stages:
 ## Results
-All BIRD-dev numbers measured on the full 1534-question set at temperature 0.7 with the
-agentic harness (execution-clustered majority vote = sample 8 candidates, group them by
-identical execution result set, return the largest cluster's query).
-| Benchmark | Metric | Score |
-|---|---|---|
-| BIRD-dev | pass@1 (avg over 8 samples) | **64.6%** EX |
-| BIRD-dev | vote-of-8 (execution-clustered) | **68.8%** EX |
-| Spider-test (2147) | pass@1 (pre-RL SFT checkpoint) | **80.3%** EX |
-Family comparison on identical seeds (pass@1 / vote@8):
 [sqrl-4b](https://huggingface.co/feyninc/sqrl-4b) 64.6 / 68.8 ·
 [sqrl-9b](https://huggingface.co/feyninc/sqrl-9b) 66.6 / 69.8 ·
 [sqrl-35b-a3b](https://huggingface.co/feyninc/sqrl-35b-a3b) 68.7 / 70.6.
 Notes from the measurement campaign:
-- Voting buys the 4B **+4.3 points** over single-pass — the largest voting gain in the
-  family (smaller models keep more sample diversity). vote-of-4 captures ~85% of that
-  gain at half the cost.
-- Base Qwen3.5-4B scores ~52% on the same harness; the pre-RL SFT checkpoint ~63.6%.
 - Vote unanimity is a strong free confidence signal: when all 8 samples agree the answer
   is right ~87% of the time; narrow/tied votes are right ~19-43% — useful for routing
   or escalation.

 ## Results
+All numbers measured with the agentic harness at temperature 0.7. BIRD-dev is the full
+1534-question set; Spider-test is 2147 questions. `pass@1` is the mean single-sample
+exec-accuracy; `vote@k` is execution-clustered majority voting over k samples (group
+candidates by identical execution result set, return the largest cluster's query).
+| Benchmark | pass@1 | vote@8 | vote@16 | vote@32 |
+|---|---|---|---|---|
+| BIRD-dev (1534) | **64.6%** | **68.8%** | 69.4% | 69.2% |
+| Spider-test (2147) | **84.0%** | **85.9%** | 85.9% | 86.1% |
+Family comparison on identical seeds (BIRD-dev pass@1 / vote@8):
 [sqrl-4b](https://huggingface.co/feyninc/sqrl-4b) 64.6 / 68.8 ·
 [sqrl-9b](https://huggingface.co/feyninc/sqrl-9b) 66.6 / 69.8 ·
 [sqrl-35b-a3b](https://huggingface.co/feyninc/sqrl-35b-a3b) 68.7 / 70.6.
 Notes from the measurement campaign:
+- Voting buys the 4B **+4.3 points** over single-pass on BIRD (the largest voting gain
+  in the family — smaller models keep more sample diversity). vote-of-4 captures ~85% of
+  that gain at half the cost.
+- **Voting saturates by k≈8–16.** Past that, more samples add nothing: BIRD peaks at
+  vote@16 (69.4) and Spider is flat from vote@8 onward (~86%). Extra test-time compute is
+  better spent on a smarter selector than on more votes.
+- Base Qwen3.5-4B scores ~52% on the same harness; the pre-RL SFT checkpoint ~63.6% on
+  BIRD (80.3% on Spider-test — RL lifts Spider single-shot to 84.0%).
 - Vote unanimity is a strong free confidence signal: when all 8 samples agree the answer
   is right ~87% of the time; narrow/tied votes are right ~19-43% — useful for routing
   or escalation.