Instructions to use cerebras/MiniMax-M2.5-REAP-139B-A10B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use cerebras/MiniMax-M2.5-REAP-139B-A10B with Transformers:

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

pipe = pipeline("text-generation", model="cerebras/MiniMax-M2.5-REAP-139B-A10B", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("cerebras/MiniMax-M2.5-REAP-139B-A10B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("cerebras/MiniMax-M2.5-REAP-139B-A10B", trust_remote_code=True)
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 cerebras/MiniMax-M2.5-REAP-139B-A10B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "cerebras/MiniMax-M2.5-REAP-139B-A10B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cerebras/MiniMax-M2.5-REAP-139B-A10B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/cerebras/MiniMax-M2.5-REAP-139B-A10B

SGLang

How to use cerebras/MiniMax-M2.5-REAP-139B-A10B 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/MiniMax-M2.5-REAP-139B-A10B" \
    --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": "cerebras/MiniMax-M2.5-REAP-139B-A10B",
		"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 "cerebras/MiniMax-M2.5-REAP-139B-A10B" \
        --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": "cerebras/MiniMax-M2.5-REAP-139B-A10B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use cerebras/MiniMax-M2.5-REAP-139B-A10B with Docker Model Runner:
```
docker model run hf.co/cerebras/MiniMax-M2.5-REAP-139B-A10B
```

Request: Step-3.5-Flash REAP variant with ~40% pruning

by rodrigomt - opened Feb 19

Discussion

rodrigomt

Feb 19

Excellent work on the MiniMax REAP versions. I'd like to know if you could create a REAP variant with a pruning rate of around 40% of the total parameters — reducing the model from its original ~196B total parameters down to approximately ~118B total parameters. This could be a great middle ground between efficiency and maximum performance, especially for users who need a lighter deployment footprint while still retaining strong reasoning capabilities.

lazarevich

Cerebras org Feb 25

Hey @rodrigomt !

we just dropped 25% and 40% pruned variants:

https://hf.co/cerebras/Step-3.5-Flash-REAP-149B-A11B
https://hf.co/cerebras/Step-3.5-Flash-REAP-121B-A11B

rodrigomt

Feb 25

Hey @lazarevich

Thank you for the quick update. I truly appreciate your time and effort dedicated to this.

LLM-Nube

Mar 1

Awesome!!! Any way to add this to the que? Qwen/Qwen3.5-397B-A17B? nvfp4 format preffered, although ggufs are awesome!

rodrigomt

Mar 3

Hey @lazarevich

Can you clear something up for me? Does the REAP version of any model work better in PyTorch, and is it more sensitive to llama.cpp quantization in GGUF format? I’ve been running into several bugs when using quantized GGUF models of any size, for example, the model getting stuck in its reasoning phase and not generating code properly.

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