mlx-community/DeepSeek-V4-Flash-2bit-DQ

Made possible by Lambda.ai ❤️

DeepSeek-V4-Flash-2bit-DQ uses a dynamic mixed-precision quantization policy. Most routed MoE expert weights are packed to 2-bit, while sensitive layers and projections remain in higher-quality 4-bit, 6-bit or 8-bit quantization. This keeps memory use much lower than the baseline 4-bit checkpoint.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/DeepSeek-V4-Flash-2bit-DQ")

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)