Kimi-K2.7-Code-mlx-DQ3_K_M-q8

mlx-community/Kimi-K2.7-Code-mlx-DQ3_K_M-q8

This model mlx-community/Kimi-K2.7-Code-mlx-DQ3_K_M-q8 was converted to MLX format from moonshotai/Kimi-K2.7-Code using mlx-lm version 0.31.3.

After the success of the first Kimi "DQ3_K_M" model, K2.5, and especially the popular K2.6 quant, this is a new update for Kimi-K2.7-Code!

This is created for people using a single Apple Mac Studio M3 Ultra with 512 GB. The 4-bit version of Kimi K2 does not fit. Using research results, we aim to get 4-bit 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.

Some points from comments of the previous K-2.6 quant:

• This model is text-only as it was made using mlx-lm.
• You need to specify --trust-remote-code. This is not something that the mlx-community added, but comes from the model provider.

You can find more similar MLX model quants for Apple Mac Studio with 512 GB at https://huggingface.co/bibproj

pip install mlx-lm

mlx_lm.generate --model mlx-community/Kimi-K2.7-Code-mlx-DQ3_K_M-q8 --trust-remote-code --prompt "Hallo"

---

What is this 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 DQ3_K_M quants?

The recipe is the same as that for the K2.5 model. Both are a bit different from that of the first Kimi "DQ3_K_M" model, which was described there. To make to the quant perform better under stress, only the expert tensors are quantized to a mix of 3-bit and 4-bit. All the other tensors are kept at 8-bit. You could say that this quant has an 8-bit "brain" and 3-bit/4-bit experts. The sizes of all three these quants are roughly the same. The 8-bit routing does reduce the tokens/second by a few %. You get a slightly slower TG, but better quality results.

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 "DQ3" 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
        if "switch_mlp.up_proj" in path:
           q_bits = 3
        if "switch_mlp.gate_proj" in path:
           q_bits = 3
        if "switch_mlp.down_proj" in path:
           q_bits = 3
           # Layers up to 5 are higher quality
           if index < 5:
              q_bits = 5
           # Every 5th layer is "medium" quality
           if (index % 5) == 0:
              q_bits = 4
        print("path:", path, "index:", index, "q_bits:", q_bits)
        return {"group_size": group_size, "bits": q_bits, "mode": mode}

Then create your DQ3_K_M quant with

mlx_lm.convert --hf-path moonshotai/Kimi-K2.7-Code --mlx-path your-model-DQ3_K_M -q --quant-predicate mixed_3_4 --trust-remote-code

NOTE*: With Kimi-K2.5, Kimi-K2.6, and Kimi-K2.7-Code you need to first dequantize the model before you can create the MLX quant. This step requires just over 2TB of additional disk space.

---

Enjoy!