AlejandroOlmedo/DeepScaleR-1.5B-Preview-6bit-mlx

About:

A fine-tuned version of Deepseek-R1-Distilled-Qwen-1.5B that surpasses the performance of OpenAI’s o1-preview with just 1.5B parameters on popular math evaluations.

Special thanks to Agentica for fine-tuning this version of Deepseek-R1-Distilled-Qwen-1.5B. More information about it can be found here:

https://huggingface.co/agentica-org/DeepScaleR-1.5B-Preview. (Base Model)

Other Types/Sizes:

Link	Type	Size	Notes
[MLX] (https://huggingface.co/AlejandroOlmedo/DeepScaleR-1.5B-Preview-8bit-mlx)	8-bit	1.90 GB	Best Quality
[MLX] (https://huggingface.co/AlejandroOlmedo/DeepScaleR-1.5B-Preview-6bit-mlx)	6-bit	1.46 GB	Better Quality
[MLX] (https://huggingface.co/AlejandroOlmedo/DeepScaleR-1.5B-Preview-4bit-mlx)	4-bit	1.01 GB	Good Quality

I simply converted it to MLX format with a quantization of 6-bits for better performance on Apple Silicon Macs (M1,M2,M3,M4 Chips).

AlejandroOlmedo/DeepScaleR-1.5B-Preview-6bit-mlx

The Model AlejandroOlmedo/DeepScaleR-1.5B-Preview-6bit-mlx was converted to MLX format from agentica-org/DeepScaleR-1.5B-Preview using mlx-lm version 0.20.5.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("AlejandroOlmedo/DeepScaleR-1.5B-Preview-6bit-mlx")

prompt="hello"

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

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