Grizzly Qwen2.5 7B Opus Distilled is an MLX-format fused fine-tune of Qwen/Qwen2.5-7B-Instruct. It was trained with LoRA adapters on Claude Opus/Fable distilled chat data, then merged back into the base model weights for direct MLX loading.

Model Description

This model is intended for local Apple Silicon experimentation with instruction-following and reasoning-oriented text generation. The training run used prompt masking so loss was applied to assistant responses rather than user prompts. The published artifact contains the fused BF16 model shards, tokenizer, chat template, and config needed for mlx-lm.

This is not an official Qwen release. It is a personal fine-tuning experiment named Grizzly, based on Qwen2.5 7B Instruct.

Training

Base model: Qwen/Qwen2.5-7B-Instruct
Training framework: mlx-lm
Fine-tune type: LoRA, fused into the published weights
LoRA rank: 16
LoRA scale: 32
Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Optimizer: AdamW
Learning rate: 1e-5
Max sequence length: 4096
Prompt masking: enabled
Final local validation loss reported by training run: 1.725

Use With MLX

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("cjnielson44/grizzly-qwen2.5-7b-opus-distilled")
response = generate(
    model,
    tokenizer,
    prompt="Explain how placebo effects work, but reason through the mechanisms carefully.",
    verbose=True,
)

Notes

This model was produced as a local fine-tuning experiment. Evaluate it for your use case before relying on outputs.