Instructions to use cerebras/DeepSeek-V3.2-REAP-508B-A37B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use cerebras/DeepSeek-V3.2-REAP-508B-A37B with Transformers:

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

pipe = pipeline("text-generation", model="cerebras/DeepSeek-V3.2-REAP-508B-A37B")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("cerebras/DeepSeek-V3.2-REAP-508B-A37B")
model = AutoModelForCausalLM.from_pretrained("cerebras/DeepSeek-V3.2-REAP-508B-A37B")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use cerebras/DeepSeek-V3.2-REAP-508B-A37B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "cerebras/DeepSeek-V3.2-REAP-508B-A37B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cerebras/DeepSeek-V3.2-REAP-508B-A37B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/cerebras/DeepSeek-V3.2-REAP-508B-A37B

SGLang

How to use cerebras/DeepSeek-V3.2-REAP-508B-A37B 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 "cerebras/DeepSeek-V3.2-REAP-508B-A37B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cerebras/DeepSeek-V3.2-REAP-508B-A37B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "cerebras/DeepSeek-V3.2-REAP-508B-A37B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cerebras/DeepSeek-V3.2-REAP-508B-A37B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use cerebras/DeepSeek-V3.2-REAP-508B-A37B with Docker Model Runner:
```
docker model run hf.co/cerebras/DeepSeek-V3.2-REAP-508B-A37B
```

FP8 versions of DeepSeek-V3.2 would awesome!

by Fernanda24 - opened Dec 17, 2025

Discussion

Fernanda24

Dec 17, 2025

•

edited Dec 17, 2025

Thanks can't wait to test but doesn't fit in 3x96gb. FP8 versions of DeepSeek-V3.2 would awesome!

Kamali-Lab

Dec 28, 2025

Pretty sure this is already FP8. DeepSeek was trained in FP8, and if this was a FP/BF16 model, the safetensors would be much larger. The safetensors here sum up to 510.57 GB. It would be at least a terabyte if it was FP16. I think what you probably need is a ~4-bit quant of this, or a quant of the 50% REAP.

Fernanda24

Dec 28, 2025

the 50% REAP fits in 4x96gb but problems loading it in sglang.. might be able to hack some changes hopefully make it work with their NSA backend. DSv31 AWQ works as it falls back to reference tilelang kernels. but this one for some reason still trying to get deepgemm even though not supported on RTX blackwell 6000 sm120 cards

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment