from mlx_lm import load, generate

model, tokenizer = load("tthoman79/Cydonia-24B-v4.3-heretic-mlx-4bit")

prompt = "hello"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True, return_dict=False,
    )

response = generate(model, tokenizer, prompt=prompt, verbose=True)

Testing notes

Tested on M3 Max, 36GB (39GB usable), via oMLX + Open WebUI.

Chat template: Confirmed intact — Mistral v7 Tekken, [SYSTEM_PROMPT]/[INST] render correctly, no leading-space issue.
Memory: Ran with TurboQuant KV cache at 8-bit.
Sampler settings that worked well: temp 0.7, min_p 0.03, top_p 0.92, repeat_penalty 1.08

Known issues

Tokenizer has the known Mistral regex bug (see mistralai/Mistral-Small-3.1-24B-Instruct-2503#84) — may cause minor tokenization edge cases. Not yet observed to cause practical issues.