Qwen3.6-27B-MXFP6-GGUF

Changed GGML TYPE to 50, latest mxfp6 build is required

This is an proof of concept/work in progress Qwen3.6-27B quantized into MXFP6.<BR> It was quantized with my <B>still experimental</B> <A HREF="https://github.com/michaelw9999/advanced-gguf-quantizer">advanced-gguf-quantizer</A> tool.<BR>

This GGUF will ONLY work on llama.cpp. The CPU only PR is posted here:

https://github.com/ggml-org/llama.cpp/pull/22671

The PR runs very slowly because that is for the initial implementation without GPU support.

You may preview the very fast POC CUDA version from my fork:

https://github.com/michaelw9999/llama.cpp/tree/mxfp6-cuda

To merge into your existing llama.cpp installation:

git remote add mxfp6 https://github.com/michaelw9999/llama.cpp
git fetch mxfp6
git merge mxfp6/mxfp6-cuda
cmake -B build -DGGML_CUDA=ON
cmake --build build -j

Or to install fresh:

git clone -b mxfp6-cuda https://github.com/michaelw9999/llama.cpp
cd llama.cpp
cmake -B build -DGGML_CUDA=ON
cmake --build build -j

NOTICE:

This is my own work and is experimental and unofficial.

The CUDA version is not part of any llama.cpp PR (yet). This is not associated with NVIDIA in anyway.

Very likely, any future MXFP6 design will not be compatible with this implementation.

While the MoE model is faster than NVFP4 on Tg, it is not yet the case on this 27B Dense model. Further optimization is still needed. The best use case will be an upcoming NVFP4/MXFP6 blend to merge the higher speeds of NVFP4 with the better quality of MXFP6 only where needed.

 Device 0: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes, VRAM: 32606 MiB
| model                          |       size |     params | backend    | ngl |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| qwen35 27B NVFP4               |  17.50 GiB |    26.90 B | CUDA       |  99 |           pp512 |       5615.34 ± 7.77 |
| qwen35 27B NVFP4               |  17.50 GiB |    26.90 B | CUDA       |  99 |           tg128 |         64.60 ± 9.77 |

  Device 0: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes, VRAM: 32606 MiB
| model                          |       size |     params | backend    | ngl |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| qwen35 27B MXFP6 - E2M3        |  20.98 GiB |    26.90 B | CUDA       |  99 |           pp512 |       3924.64 ± 4.79 |
| qwen35 27B MXFP6 - E2M3        |  20.98 GiB |    26.90 B | CUDA       |  99 |           tg128 |         41.15 ± 3.57 |

Feedback is both requested and encouraged so I can make further improvements into future llama.cpp PRs.

The NVFP4/MXFP6 quantizer is still being improved and will be posted in the future. Please let me know if you want to see a specific model turned into MXFP6.

I will create an MTP enabled version soon.

The quality improvement of MXFP6 vs NVFP4 is quite striking! The full kld against the BF16 model is below.

MXFP6 KLD vs BF16:

====== Perplexity statistics ======
Mean PPL(Q)                   :   6.918624 ±   0.045510
Mean PPL(base)                :   6.900856 ±   0.045374
Cor(ln(PPL(Q)), ln(PPL(base))):  99.52%
Mean ln(PPL(Q)/PPL(base))     :   0.002571 ±   0.000645
Mean PPL(Q)/PPL(base)         :   1.002575 ±   0.000646
Mean PPL(Q)-PPL(base)         :   0.017768 ±   0.004457

====== KL divergence statistics ======
Mean    KLD:   0.018925 ±   0.000647
Maximum KLD:  28.370348
99.9%   KLD:   2.194927
99.0%   KLD:   0.141343
95.0%   KLD:   0.038436
90.0%   KLD:   0.022466
Median  KLD:   0.004691
10.0%   KLD:   0.000123
 5.0%   KLD:   0.000033
 1.0%   KLD:   0.000004
 0.1%   KLD:  -0.000001
Minimum KLD:  -0.000054

====== Token probability statistics ======
Mean    Δp: -0.102 ± 0.010 %
Maximum Δp: 99.909%
99.9%   Δp: 27.867%
99.0%   Δp:  8.482%
95.0%   Δp:  3.660%
90.0%   Δp:  2.084%
75.0%   Δp:  0.405%
Median  Δp: -0.001%
25.0%   Δp: -0.540%
10.0%   Δp: -2.405%
 5.0%   Δp: -4.104%
 1.0%   Δp: -9.415%
 0.1%   Δp: -29.850%
Minimum Δp: -99.768%
RMS Δp    :  3.662 ± 0.060 %
Same top p: 95.023 ± 0.057 %

and for comparison, NVFP4:

====== Perplexity statistics ======
Mean PPL(Q)                   :   7.321749 ±   0.049291
Mean PPL(base)                :   6.900856 ±   0.045374
Cor(ln(PPL(Q)), ln(PPL(base))):  98.00%
Mean ln(PPL(Q)/PPL(base))     :   0.059204 ±   0.001341
Mean PPL(Q)/PPL(base)         :   1.060991 ±   0.001423
Mean PPL(Q)-PPL(base)         :   0.420893 ±   0.010245

====== KL divergence statistics ======
Mean    KLD:   0.079692 ±   0.001052
Maximum KLD:  24.928333
99.9%   KLD:   5.620802
99.0%   KLD:   0.820752
95.0%   KLD:   0.239744
90.0%   KLD:   0.137482
Median  KLD:   0.027444
10.0%   KLD:   0.000694
 5.0%   KLD:   0.000195
 1.0%   KLD:   0.000026
 0.1%   KLD:   0.000004
Minimum KLD:  -0.000092

====== Token probability statistics ======
Mean    Δp: -0.808 ± 0.020 %
Maximum Δp: 99.897%
99.9%   Δp: 49.456%
99.0%   Δp: 17.496%
95.0%   Δp:  7.536%
90.0%   Δp:  4.210%
75.0%   Δp:  0.701%
Median  Δp: -0.019%
25.0%   Δp: -1.609%
10.0%   Δp: -6.312%
 5.0%   Δp: -10.856%
 1.0%   Δp: -28.085%
 0.1%   Δp: -73.216%
Minimum Δp: -99.926%
RMS Δp    :  7.764 ± 0.063 %
Same top p: 89.020 ± 0.081 %