mlx-community/DeepSeek-V3.1-mlx-DQ5_K_M

This model mlx-community/DeepSeek-V3.1-mlx-DQ5_K_M was converted to MLX format from deepseek-ai/DeepSeek-V3.1 using mlx-lm version 0.26.3.

This is created for people using a single Apple Mac Studio M3 Ultra with 512 GB. With 512 GB, we can do better than the 4-bit version of DeepSeek V3.1. Using research results, we aim to get better than 5-bit performance using smarter quantization. We aim to not have the quant so large that it leaves no memory for a useful context window.

pip install mlx-lm

mlx_lm.generate --model mlx-community/DeepSeek-V3.1-mlx-DQ5_K_M --temp 1.3 --min-p 0.01 --max-tokens 4096 --prompt "Hallo"

The temperature of 1.3 is DeepSeek's recommendation for translations. For coding, you should probably use a temperature of 0.6 or lower.

---

What is this DQ5_K_M?

In the Arxiv paper Quantitative Analysis of Performance Drop in DeepSeek Model Quantization the authors write,

We further propose DQ3_K_M, a dynamic 3-bit quantization method that significantly outperforms traditional Q3_K_M variant on various benchmarks, which is also comparable with 4-bit quantization (Q4_K_M) approach in most tasks.

and

dynamic 3-bit quantization method (DQ3_K_M) that outperforms the 3-bit quantization implementation in llama.cpp and achieves performance comparable to 4-bit quantization across multiple benchmarks.

The resulting multi-bitwidth quantization has been well tested and documented.

In this case we did not want a improved 3-bit quant, but rather the best possible "5-bit" quant. We therefore modified the DQ3_K_M quantization by replacing 3-bit by 5-bit, 4-bit by 6-bit, and 6-bit by 8-bit to create a new DQ5_K_M quant. This produces a quantization of 5.638 bpw (bits per weight).

---

How can you create your own DQ5_K_M quants?

In the convert.py file of mlx-lm on your system ( you can see the original code here ), replace the code inside def mixed_quant_predicate() with something like

        index = (
            int(path.split(".")[layer_location])
            if len(path.split(".")) > layer_location
            else 0
        )
        # Build a mixed quant like "DQ5" similar to the "DQ3" of Arxiv paper https://arxiv.org/abs/2505.02390
        #    Quantitative Analysis of Performance Drop in DeepSeek Model Quantization
        q_bits = 6       
        if "lm_head" in path:
           q_bits = 8
        #if "tokens" in path:
        #   q_bits = 6
        if "attn.kv" in path:
           q_bits = 8
        #if "o_proj" in path:
        #   q_bits = 6
        #if "attn.q" in path:
        #   q_bits = 6
        # For all "mlp" and "shared experts"
        if "down_proj" in path:
           q_bits = 8
        #if "up_proj" in path:
        #   q_bits = 6
        #if "gate_proj" in path:
        #   q_bits = 6
        # For "switch experts"
        if "switch_mlp.up_proj" in path:
           q_bits = 5
        if "switch_mlp.gate_proj" in path:
           q_bits = 5
        if "switch_mlp.down_proj" in path:
           q_bits = 5
           # Blocks up to 5 are higher quality
           if index < 5:
              q_bits = 8
           # Every 5th block is "medium" quality
           if (index % 5) == 0:
              q_bits = 6
        #print("path:", path, "index:", index, "q_bits:", q_bits)
        return {"group_size": group_size, "bits": q_bits}

Should you wish to squeeze more out of your quant, and you do not need to use a larger context window, you can change the last part of the above code to

        if "switch_mlp.down_proj" in path:
           q_bits = 6
           # Blocks up to 5 are higher quality
           if index < 5:
              q_bits = 8
        #print("path:", path, "index:", index, "q_bits:", q_bits)
        return {"group_size": group_size, "bits": q_bits}

Then create your DQ5_K_M quant with

mlx_lm.convert --hf-path deepseek-ai/DeepSeek-V3.1 --mlx-path your-model-DQ5_K_M -q --quant-predicate mixed_3_4

---

Enjoy!