gghfez's picture
Create README.md
667f29b verified
metadata
base_model:
  - mistralai/Mistral-Small-3.1-24B-Instruct-2503

Text-Only EXL2 Quant of mistralai/Mistral-Small-3.1-24B-Instruct-2503

The following changes were made:

  1. Vision encoder removed
  2. Architecture changed to that of mistralai/Mistral-Small-24B-Instruct-2501
  3. Chat template in tokenizer_config.json was modified (see below).

I was having trouble with the timestamp at the beginning of the system prompt and removed it from tokenizer_config.json.

NOTE Tensor Parallel is not implemented in exllamav2 for both mistralai/Mistral-Small-3.1-24B-Instruct-2503 and mistralai/Mistral-Small-24B-Instruct-2501.