---
license: apache-2.0
base_model: ProCreations/grug-27b
tags:
- grug
- mtp
- speculative-decoding
- reasoning
language:
- en
---

# grug-27b-mtp

grug think fast. now grug PREDICT own thought before think it. free speedup.

this is [grug-27b](https://huggingface.co/ProCreations/grug-27b) (v2.1) with a
**trained MTP head** for multi-token prediction / speculative decoding. base
Qwen3.6-27B ship MTP head; ordinary grug-27b drop it. grug graft head back AND
retrain it on grug data so draft match grug's terse output style.

## number

| draft quality (t+2 top-1 agreement on grug outputs) | % |
|---|---:|
| base Qwen MTP head on grug backbone (grafted, untrained) | 81.7 |
| after grug-training the head | **89.3** |

higher agreement = more accepted drafts = more free speedup. head trained on
3M token of grug v2.1 data, backbone completely frozen (identical quality to
grug-27b - same tensors bit-for-bit).

## how run

vLLM (>= 0.19):
```bash
vllm serve ProCreations/grug-27b-mtp --max-model-len 32768 \
  --reasoning-parser deepseek_r1 \
  --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'
```

gguf with MTP tensors: [grug-27b-mtp-gguf](https://huggingface.co/ProCreations/grug-27b-mtp-gguf)
(needs llama.cpp build with qwen3_5 MTP speculative support).

no MTP wanted? use plain [grug-27b](https://huggingface.co/ProCreations/grug-27b) -
same brain, smaller checkpoint. grug made by ProCreations.