mlx-community/LongCat-Flash-Chat-mlx-DQ6_K_M

This model mlx-community/LongCat-Flash-Chat-mlx-DQ6_K_M was converted to MLX format from meituan-longcat/LongCat-Flash-Chat using mlx-lm version 0.28.1.

This is created for people using a single Apple Mac Studio M3 Ultra with 512 GB. The 8-bit version of Ring 1T does not fit. Using research results, we aim to get almost-8bit performance from a slightly smaller and smarter quantization. It should also not be so large that it leaves no memory for a useful context window.

pip install mlx-lm

mlx_lm.generate --model mlx-community/LongCat-Flash-Chat-mlx-DQ6_K_M --temp 0.7 --max-tokens 4096 --prompt "Hallo"

---

What is the DQ6_K_M quant about? It comes from a paper on DQ3_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.

---

How can you create your own DQ6_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 similar to the "DQ3" of Arxiv paper https://arxiv.org/abs/2505.02390
        #    Quantitative Analysis of Performance Drop in DeepSeek Model Quantization
        q_bits = 8
        # For "switch experts"
        if "switch_mlp" in path:
           q_bits = 6
        if "switch_mlp.down_proj" in path:
           # 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 = 8
        print("path:", path, "index:", index, "q_bits:", q_bits)
        return {"group_size": group_size, "bits": q_bits}

Then create your DQ6_K_M quant with

mlx_lm.convert --hf-path meituan-longcat/LongCat-Flash-Chat --mlx-path your-model-DQ6_K_M -q --quant-predicate mixed_3_4

---

Enjoy!