These are the MTP tensors for Qwen3.6-27B, quantized at 4 bpw.

Qwen3.5/3.6 checkpoints quantized with exllamav3 v0.0.41+ will automatically contain MTP 
layers. For quantized checkpoints from earlier versions, place this file in the model 
directory to enable MTP drafting.

Quick benchmark (RTX 6000 Pro, 4.00 bpw target model):

| Category                |   Baseline |      MTP (greedy) |
|:------------------------|-----------:|------------------:|
| Agentic, code           |  63.88 t/s | 180.51 t/s, 2.83x |
| Agentic, curl           |  61.42 t/s | 154.40 t/s, 2.51x |
| Coding                  |  68.17 t/s | 187.97 t/s, 2.76x |
| Creative                |  68.29 t/s | 138.22 t/s, 2.02x |
| Creative (reasoning)    |  68.01 t/s | 135.00 t/s, 1.98x |
| Translation             |  66.88 t/s | 139.41 t/s, 2.08x |
| Translation (reasoning) |  66.79 t/s | 162.90 t/s, 2.44x |
| Trivial repetition      |  68.87 t/s | 237.59 t/s, 3.45x |