1from mlx_lm import load, generate
2
3model, tokenizer = load("salohcin714/granite-4.1-3b-mxfp8-mlx")
4messages = [{"role": "user", "content": "Hello"}]
5prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
6text = generate(model, tokenizer, prompt=prompt, verbose=True)
Weights converted to MLX safetensors layout and quantized (MXFP8 microscaling floating-point quantization, via round-to-nearest, no calibration). Redundant tied lm_head.weight dropped where the model ties input/output embeddings. No fine-tuning; no added training data.
Licensed under
Apache 2.0. Original weights by the Granite Team, IBM. See the
upstream model card and the included
LICENSE file for the full text.
This repository is not affiliated with or endorsed by IBM. "Granite" is an IBM trademark, used here descriptively to identify the origin of the base model. IBM's published benchmarks describe the original weights, not this quantized/converted artifact, and must not be read as claims about this repo.
