Modified variant: both the forced identity and the default system prompt have been completely removed from chat_template.jinja. A system block is emitted only when the caller provides one. The weights are unchanged and identical to the same-quantization base variant.

Usage (mlx-vlm)

uv add mlx-vlm

python

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("ToPo-ToPo/Qwythos-9B-Claude-Mythos-5-1M-mlx-8bit-noident")
messages = [{"role": "user", "content": "Hello"}]
prompt = apply_chat_template(processor, model.config, messages)
print(generate(model, processor, prompt, max_tokens=256))

This repository is a derivative work distributed under Apache-2.0. The original attribution and license are inherited as noted above.