dicksondickson/Carbon-X-oQ8e-bf16-MLX

This model is a quantization of:
https://huggingface.co/HuggingFaceBio/Carbon-X

This model was quantized using oQ (oMLX 0.7.3.dev2) imatrix, mixed-precision quantization with Better Low-Bit oQ.

Important tensors are left at bf16 which is for Apple M3 chips and later.

Run the model using oMLX:
https://github.com/jundot/omlx

The tokenizer has been patch to run with oMLX and runs completely offline.

Please use the following settings in oMLX:

Model Settings → Advanced → Trust Remote Code → on

| Temperature | 0.6
| Top-p | 0.95
| Top-k | 20
| Thinking | Off (enable_thinking=False)
| Max tokens | Cap to what you need (e.g. 1024–4096) | Long DNA generation is slow, and its quality drops the longer it runs | Context | Up to 40,960 tokens (~245 kbp of DNA)