Instructions to use superwhisper/s1-mini with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use superwhisper/s1-mini with Transformers:

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

pipe = pipeline("text-generation", model="superwhisper/s1-mini")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("superwhisper/s1-mini")
model = AutoModelForCausalLM.from_pretrained("superwhisper/s1-mini", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use superwhisper/s1-mini with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "superwhisper/s1-mini"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "superwhisper/s1-mini",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/superwhisper/s1-mini

SGLang

How to use superwhisper/s1-mini 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 "superwhisper/s1-mini" \
    --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": "superwhisper/s1-mini",
		"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 "superwhisper/s1-mini" \
        --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": "superwhisper/s1-mini",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use superwhisper/s1-mini with Docker Model Runner:
```
docker model run hf.co/superwhisper/s1-mini
```

Softmaximalist commited on 17 days ago

Commit

3069a5d

1 Parent(s): 1443829

Add app integration instructions

Browse files

Files changed (1) hide show

README.md +36 -7

README.md CHANGED Viewed

@@ -249,8 +249,10 @@ in its training data.
 > common way to get a blank result from this model.
 The flag makes the template emit an empty think block before the assistant
-turn. That is the exact prefix the model saw during training. If you build
-prompts by hand instead of using the template, the literal string is:
 ```
 <|im_start|>system
@@ -324,11 +326,7 @@ with the empty think block shown above, or you will get blank output.
    leaving it at 1024.
 4. **Chunk long transcripts at sentence boundaries.** The model is built for
    dictation-length input; keep single passes under roughly 1,000 tokens.
-5. **Guard against stuck ASR.** If the upstream ASR gets stuck and repeats a
-   phrase many times, greedy decoding can mirror the loop until it hits
-   `max_new_tokens`. A `repetition_penalty` of 1.1 terminates cleanly without
-   affecting normal inputs.
-6. **Expect an empty string sometimes.** Filler-only input returns nothing,
    and your pipeline should treat that as a valid result rather than a
    failure.
@@ -339,6 +337,37 @@ output and synthetic stress sets for numbers, self-corrections, lists, email
 and adversarial inputs: 94.4% token accuracy, measured greedy on the Q4_K_M
 GGUF build. The BF16 weights here should do at least as well.
 ## License
 s1-mini is released under Apache 2.0, which it inherits from Qwen3-0.6B.

 > common way to get a blank result from this model.
 The flag makes the template emit an empty think block before the assistant
+turn. That is the exact prefix the model saw during training. The template
+ships as `chat_template.jinja` in this repo, so `apply_chat_template` picks it
+up with no extra configuration. If you build prompts by hand instead, the
+literal string is:
 ```
 <|im_start|>system
    leaving it at 1024.
 4. **Chunk long transcripts at sentence boundaries.** The model is built for
    dictation-length input; keep single passes under roughly 1,000 tokens.
+5. **Expect an empty string sometimes.** Filler-only input returns nothing,
    and your pipeline should treat that as a valid result rather than a
    failure.
 and adversarial inputs: 94.4% token accuracy, measured greedy on the Q4_K_M
 GGUF build. The BF16 weights here should do at least as well.
+## Using s1-mini in your own app
+s1-mini is Apache 2.0, the same license it inherits from Qwen3-0.6B, so it can
+be embedded in open-source and commercial software alike: dictation apps,
+meeting-notes tools, live captioning, voice-driven editors, or any pipeline
+that has to turn raw ASR output into text a person will read.
+It is a post-processing stage rather than a standalone system:
+```
+audio ──▶ ASR (Whisper, Parakeet, …) ──▶ s1-mini ──▶ clean text
+```
+The ASR's raw transcript becomes the transcript line, your app's settings
+choose the three control-line values, and the model returns text ready to
+display. At 0.6B it is small enough to ship on-device, and the
+[GGUF builds](https://huggingface.co/superwhisper/s1-mini-GGUF) exist for
+exactly that.
+Nothing about the model is superwhisper-specific. The two things to get right
+in any integration are the input format documented above and the thinking
+flag; nearly every integration bug traces back to one of those.
+> [!IMPORTANT]
+> Read the [LICENSE](LICENSE) before you ship. Apache 2.0 is permissive but
+> not obligation-free: you must retain the license text and attribution
+> notices, and state significant changes if you redistribute a modified
+> version. If you are bundling s1-mini into a commercial dictation app or
+> redistributing the weights yourself, confirm the terms cover your case
+> rather than assuming they do.
 ## License
 s1-mini is released under Apache 2.0, which it inherits from Qwen3-0.6B.