File size: 4,157 Bytes
87acb41
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bc0af79
87acb41
bc0af79
87acb41
 
 
 
 
 
 
 
bc0af79
87acb41
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
07be91a
 
87acb41
 
 
 
 
 
 
 
 
 
 
 
 
07be91a
 
87acb41
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
---
license: other
license_name: modified-mit
license_link: https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905/blob/main/LICENSE
library_name: mlx
tags:
- mlx
pipeline_tag: text-generation
base_model: moonshotai/Kimi-K2-Instruct-0905
---

. . .

# UPLOADING FILES ...

---

# mlx-community/Kimi-K2-Instruct-0905-mlx-DQ3_K_M

This model [mlx-community/Kimi-K2-Instruct-0905-mlx-DQ3_K_M](https://huggingface.co/mlx-community/Kimi-K2-Instruct-0905-mlx-DQ3_K_M) was
converted to MLX format from [moonshotai/Kimi-K2-Instruct-0905](https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905)
using mlx-lm version **0.26.3**.

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.

```bash
pip install mlx-lm

mlx_lm.generate --model mlx-community/Kimi-K2-Instruct-0905-mlx-DQ3_K_M --temp 0.6 --min-p 0.01 --max-tokens 4096 --trust-remote-code --prompt "Hallo"
```

---

## What is this DQ3_K_M?

In the Arxiv paper [Quantitative Analysis of Performance Drop in DeepSeek Model Quantization](https://arxiv.org/abs/2505.02390) 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?

In the `convert.py` file of mlx-lm on your system ( [you can see the original code here](https://github.com/ml-explore/mlx-lm/blob/main/mlx_lm/convert.py) ), replace the code inside `def mixed_quant_predicate()` with something like

```python
        index = (
            int(path.split(".")[layer_location])
            if len(path.split(".")) > layer_location
            else 0
        )
        # Build a mixed quant like "DQ3" of Arxiv paper https://arxiv.org/abs/2505.02390
        #    Quantitative Analysis of Performance Drop in DeepSeek Model Quantization
        q_bits = 4       
        if "lm_head" in path:
           q_bits = 6
        #if "tokens" in path:
        #   q_bits = 4
        if "attn.kv" in path:
           q_bits = 6
        #if "o_proj" in path:
        #   q_bits = 4
        #if "attn.q" in path:
        #   q_bits = 4
        # For all "mlp" and "shared experts"
        if "down_proj" in path:
           q_bits = 6
        #if "up_proj" in path:
        #   q_bits = 4
        #if "gate_proj" in path:
        #   q_bits = 4
        # For "switch experts"
        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
           # Blocks up to 5 are higher quality
           if index < 5:
              q_bits = 6
           # Every 5th block 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}
```

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

```python
        if "switch_mlp.down_proj" in path:
           q_bits = 4
           # Blocks up to 5 are higher quality
           if index < 5:
              q_bits = 6
        #print("path:", path, "index:", index, "q_bits:", q_bits)
        return {"group_size": group_size, "bits": q_bits}
```

Then create your DQ3_K_M quant with

```bash
mlx_lm.convert --hf-path moonshotai/Kimi-K2-Instruct-0905 --mlx-path your-model-DQ3_K_M -q --quant-predicate mixed_3_4 --trust-remote-code
```

---

Enjoy!