Instructions to use ProjectAstroAI/Sonke-1.0-350M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ProjectAstroAI/Sonke-1.0-350M with Transformers:

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

pipe = pipeline("text-generation", model="ProjectAstroAI/Sonke-1.0-350M", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("ProjectAstroAI/Sonke-1.0-350M", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ProjectAstroAI/Sonke-1.0-350M with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ProjectAstroAI/Sonke-1.0-350M"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ProjectAstroAI/Sonke-1.0-350M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ProjectAstroAI/Sonke-1.0-350M

SGLang

How to use ProjectAstroAI/Sonke-1.0-350M 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 "ProjectAstroAI/Sonke-1.0-350M" \
    --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": "ProjectAstroAI/Sonke-1.0-350M",
		"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 "ProjectAstroAI/Sonke-1.0-350M" \
        --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": "ProjectAstroAI/Sonke-1.0-350M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ProjectAstroAI/Sonke-1.0-350M with Docker Model Runner:
```
docker model run hf.co/ProjectAstroAI/Sonke-1.0-350M
```

Sönke-1.0-350M

A 350M model I'm training from scratch on my own GPU. It can chat, do basic math and write small bits of code, and it can think before it answers. This is step 58,666 (22.8B tokens), so it's not done yet. Newer checkpoints will be uploaded as training goes.

How to use it

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "ProjectAstroAI/Sonke-1.0-350M"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True, dtype="bfloat16").to("cuda")

messages = [{"role": "user", "content": "Explain photosynthesis in two sentences."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.9, do_sample=True)
print(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Remove enable_thinking=False if you want it to think first. It needs trust_remote_code=True because the architecture is my own. Batches need prompts of the same length for now.

Evals

Model	MMLU	HellaSwag	ARC-C	GSM8K	HumanEval
Sönke-1.0-350M	28.6	42.0	28.8	12.3	12.2
SmolLM2-360M	25.0	54.5	38.4	3.2	12.2
Qwen2.5-0.5B	47.5	52.1	35.6	41.6	30.5
Qwen2.5-1.5B	60.9	67.9	54.7	68.5	37.2

Sönke: MMLU 5-shot, HellaSwag and ARC-C 0-shot, GSM8K 0-shot without thinking, HumanEval pass@1. SmolLM2's MMLU, ARC-C and HumanEval were run the same way; everything else is from the models' own cards, with different shot counts, so small gaps don't mean much.

Limitations

It's small and still training. It gets facts wrong, loses track in long chats, and sometimes messes up formatting. Don't use it for anything that matters.

License

Apache 2.0. Part of the training data was generated with DeepSeek models, so use of this model is also subject to the use restrictions in LICENSE-DEEPSEEK-MODEL (paragraph 5 and Attachment A). Dataset credits are in NOTICE.

Downloads last month: -

Safetensors

Model size

0.3B params

Tensor type

F32

Datasets used to train ProjectAstroAI/Sonke-1.0-350M

Collection including ProjectAstroAI/Sonke-1.0-350M

Sönke

Collection

A series of 350M parameter AI language models. • 1 item • Updated about 2 hours ago