Instructions to use tomasmcm/sky-t1-coder-32b-flash with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use tomasmcm/sky-t1-coder-32b-flash with Transformers:

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

pipe = pipeline("text-generation", model="tomasmcm/sky-t1-coder-32b-flash")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("tomasmcm/sky-t1-coder-32b-flash")
model = AutoModelForCausalLM.from_pretrained("tomasmcm/sky-t1-coder-32b-flash")
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

vLLM

How to use tomasmcm/sky-t1-coder-32b-flash with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "tomasmcm/sky-t1-coder-32b-flash"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tomasmcm/sky-t1-coder-32b-flash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/tomasmcm/sky-t1-coder-32b-flash

SGLang

How to use tomasmcm/sky-t1-coder-32b-flash 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 "tomasmcm/sky-t1-coder-32b-flash" \
    --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": "tomasmcm/sky-t1-coder-32b-flash",
		"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 "tomasmcm/sky-t1-coder-32b-flash" \
        --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": "tomasmcm/sky-t1-coder-32b-flash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use tomasmcm/sky-t1-coder-32b-flash with Docker Model Runner:
```
docker model run hf.co/tomasmcm/sky-t1-coder-32b-flash
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

tomasmcm/sky-t1-coder-32b-flash

This is a merge of pre-trained language models created using mergekit.

I wanted to see if it would be possible to improve on FuseO1-DeepSeekR1-Qwen2.5-Coder-32B-Preview and CoderO1-DeepSeekR1-Coder-32B-Preview by using Sky-T1-32B-Flash as the reasoning model that is merged with Qwen2.5-Coder-32B-Instruct instead of DeepSeek-R1-Distill-Qwen-32B. The idea is to have a strong coder model that can reason but without very long reasoning chains (hence using the Flash model).

GGUF files available at mradermacher/sky-t1-coder-32b-flash-GGUF (thank you!)

LLM Leaderboard (15 Mar 2025)

Rank	Model	Average	IFEval	BBH	MATH	GPQA	MUSR	MMLU-PRO
22	Qwen/Qwen2.5-32B-Instruct	46.60%	83.46%	56.49%	62.54%	11.74%	13.50%	51.85%
39	tomasmcm/sky-t1-coder-32b-flash	44.87%	77.80%	55.47%	54.23%	15.77%	12.81%	53.13%
147	huihuiai/QwQ-32B-Coder-Fusion-9010	41.58%	57.78%	53.02%	53.17%	14.88%	19.52%	51.11%
276	Qwen/Qwen2.5-Coder-32B-Instruct	39.89%	72.65%	52.27%	49.55%	13.20%	13.72%	37.92%
717	Qwen/QwQ-32B-Preview	34.12%	40.35%	53.39%	44.94%	4.25%	9.81%	51.98%
2100	deepseek-ai/DeepSeek-R1-Distill-Qwen-32B	22.96%	41.86%	17.15%	17.07%	4.59%	16.14%	40.96%
3574	Qwen/QwQ-32B	12.21%	39.77%	2.87%	16.09%	1.34%	11.05%	2.18%

Merge Details

Merge Method

This model was merged using the SCE merge method using Qwen/Qwen2.5-Coder-32B as a base.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

models:
  # Pivot model
  - model: Qwen/Qwen2.5-Coder-32B
  # Target models
  - model: Qwen/Qwen2.5-Coder-32B-Instruct
  - model: NovaSky-AI/Sky-T1-32B-Flash
merge_method: sce
base_model: Qwen/Qwen2.5-Coder-32B
parameters:
  select_topk: 1.0
dtype: bfloat16