GiaHuy/Darwin-36B-Opus-mlx-text-only-8bit

Darwin-36B-Opus MLX Text-Only 8-bit

This repository contains a text-only 8-bit MLX conversion of FINAL-Bench/Darwin-36B-Opus.

The model was converted with mlx-lm and is intended for efficient inference on Apple Silicon.

Model Details

• Original model: FINAL-Bench/Darwin-36B-Opus
• Format: MLX
• Quantization: 8-bit
• Modality: Text-only
• Runtime: Apple Silicon
• Recommended server: cubist38/mlx-openai-server

Run with mlx-openai-server

This model is designed to be served through my open-source OpenAI-compatible MLX server:

cubist38/mlx-openai-server

Install the server:

pip install mlx-openai-server

Then launch the model:

mlx-openai-server launch \
  --model-path Darwin-36B-Opus-mlx-text-only-8bit \
  --reasoning-parser qwen3_moe \
  --tool-call-parser qwen3_coder \
  --debug \
  --served-model-name Darwin-36B-Opus

The server exposes an OpenAI-compatible API, making it easy to use with existing OpenAI SDKs, agents, and tools.

Example: OpenAI-Compatible Python Client

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="not-needed",
)

response = client.chat.completions.create(
    model="Darwin-36B-Opus",
    messages=[
        {
            "role": "user",
            "content": "Explain evolutionary model merging in simple terms.",
        }
    ],
    temperature=0.7,
    max_tokens=512,
)

print(response.choices[0].message.content)

Example: curl

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer not-needed" \
  -d '{
    "model": "Darwin-36B-Opus",
    "messages": [
      {
        "role": "user",
        "content": "What makes Darwin-36B-Opus interesting?"
      }
    ],
    "temperature": 0.7,
    "max_tokens": 512
  }'

Notes

• This is a text-only MLX conversion.
• This is an 8-bit quantized version, so outputs may differ from the original checkpoint.
• The recommended way to serve this model is through mlx-openai-server.
• The launch command uses:
  • --reasoning-parser qwen3_moe
  • --tool-call-parser qwen3_coder
  • --served-model-name Darwin-36B-Opus

Attribution

All credit for the original model goes to FINAL-Bench/Darwin-36B-Opus.

This repository provides only an MLX text-only 8-bit conversion for Apple Silicon users.

License

Please refer to the original model repository for licensing and usage terms.