granite-4.1-8B-TNG-Coder-V11-3000-Heretic-BF16/README.md

license, base_model, pipeline_tag, library_name, tags

license	base_model	pipeline_tag	library_name	tags
apache-2.0	ibm-granite/granite-4.1-8b treadon/granite-4.1-8b-Abliterated-AND-Disinhibited nightmedia/granite-4.1-8B-TNG-Coder-V11-3000-Heretic	text-generation	transformers	finetune granite-4.1 reasoning thinking emergent mlx

granite-4.1-8B-TNG-Coder-V11-3000-Heretic-BF16

This model was trained on a base of treadon/granite-4.1-8b-Abliterated-AND-Disinhibited.

The training was exclusively in backend development, and took place on DS9, with Spock and Data for Haskell, Worf for Golang, Odo for Python, assisted by Garak, Quark, and Q.

The teacher was a Qwen3.6-35B-A3B with special tuning, profiled to respond in character, think, and serve humour as needed. A second round of training added a rich set of Claude traces, mixed in with lessons already learned.

It has thinking mode by default, unlike the parent model.

         arc   arc/e boolq hswag obkqa piqa  wino
bf16     0.561,0.756,0.862,0.747,0.438,0.805,0.711
q8-hi    0.562,0.755,0.861,0.746,0.428,0.803,0.703
qx86-hi  0.565,0.756,0.863,0.747,0.436,0.804,0.712

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("granite-4.1-8B-TNG-Coder-V11-3000-Heretic-BF16")

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)