This is the second release of the NVFP4 version of Jackrong's Qwopus3.6-27B-v2-MTP-GGUF.
9-June-2026: Fixed MTP heads to NVFP4; performance with MTP is much improved (108tk/s tg)
Please note, I am not affiliated; this is my own quantization effort made with my experimental work-in-progress advanced-gguf-quantizer.
More evaluatlions will be underway, this page will be updated when those are complete.
Feedback on how to improve the quantizer/this quantization is appreciated.
For improved performance and quality, try my llama.cpp NVFP4-Repack with MXFP6 from:
https://github.com/michaelw9999/llama.cpp/tree/nvfp4repack_mxfp6_cuda
These branches are updated regularly.
NVFP4 repack preloads all tensors into a CUDA tile to boost speed. It is a tiny bit slower on first load, then provides ~10% prefill boost with a small reduction in token gen seen on larger models, and an increase on smaller models.
However, it also enables NVFP4 input scale, which boosts model correctness.
Initial performance results:
llama-bench on 5090:
qwen35 27B NVFP4 | 15.14 GiB |27.32 B | CUDA | pp512 | 5958.06 ± 6.46 |
qwen35 27B NVFP4 | 15.14 GiB |27.32 B | CUDA | tg128 | 73.72 ± 0.08 |
Perplexity/kld results against wiki2 test:
====== Perplexity statistics ======
Mean PPL(Q) : 6.959259 ± 0.045420
Mean PPL(base) : 6.694312 ± 0.042968
Cor(ln(PPL(Q)), ln(PPL(base))): 98.93%
Mean ln(PPL(Q)/PPL(base)) : 0.038815 ± 0.000953
Mean PPL(Q)/PPL(base) : 1.039578 ± 0.000991
Mean PPL(Q)-PPL(base) : 0.264947 ± 0.006915
====== KL divergence statistics ======
Mean KLD: 0.045311 ± 0.000650
Maximum KLD: 19.236860
99.9% KLD: 2.298393
99.0% KLD: 0.423422
95.0% KLD: 0.135200
90.0% KLD: 0.081522
Median KLD: 0.017625
10.0% KLD: 0.000521
5.0% KLD: 0.000152
1.0% KLD: 0.000024
0.1% KLD: 0.000004
Minimum KLD: -0.000062
====== Token probability statistics ======
Mean Δp: -0.450 ± 0.016 %
Maximum Δp: 99.955%
99.9% Δp: 32.792%
99.0% Δp: 13.999%
95.0% Δp: 6.515%
90.0% Δp: 3.788%
75.0% Δp: 0.717%
Median Δp: -0.012%
25.0% Δp: -1.153%
10.0% Δp: -4.684%
5.0% Δp: -8.120%
1.0% Δp: -20.878%
0.1% Δp: -55.713%
Minimum Δp: -99.221%
RMS Δp : 6.022 ± 0.055 %
Same top p: 91.085 ± 0.074 %