Bond v1 — first-person character voice (MLX 4-bit)

MLX 4-bit quantization (affine, group size 64) of darthcrawl/bond, for Apple Silicon (mlx-lm / LM Studio). ~12 GB.

Bond is a first-person character-voice fine-tune of Mistral-Small-3.1-24B-Instruct: it holds a committed in-character first-person voice instead of leaking back to helpful-assistant mode. Uncensored, explicit-capable. Eval, training, and the design goal are on the base model card.

This quant

Smallest. Comfortable on a 24GB Mac (tight on 16GB). Minor quality cost vs bf16 — fine for most use.

Other MLX sizes: 6-bit, 8-bit. (4-bit smallest, 6-bit near-lossless, 8-bit lossless.)

Use

from mlx_lm import load, generate

model, tok = load("darthcrawl/bond-mlx-4bit")
prompt = tok.apply_chat_template(
    [{"role": "user", "content": "Tell me about the worst night of your life."}],
    add_generation_prompt=True,
)
print(generate(model, tok, prompt=prompt, max_tokens=512, verbose=True))