grug think fast. now grug PREDICT own thought before think it. free speedup.
this is 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.
| 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).
vLLM (>= 0.19):
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 (needs llama.cpp build with qwen3_5 MTP speculative support).
no MTP wanted? use plain grug-27b - same brain, smaller checkpoint. grug made by ProCreations.
model-00014-of-00015.safetensors 4.0 GBmodel-00009-of-00015.safetensors 4.0 GBmodel-00003-of-00015.safetensors 4.0 GBmodel-00006-of-00015.safetensors 4.0 GBmodel-00012-of-00015.safetensors 4.0 GBmodel-00013-of-00015.safetensors 4.0 GBmodel-00010-of-00015.safetensors 4.0 GBmodel-00004-of-00015.safetensors 4.0 GBmodel-00007-of-00015.safetensors 3.9 GBmodel-00002-of-00015.safetensors 3.9 GBmodel-00005-of-00015.safetensors 3.9 GBmodel-00011-of-00015.safetensors 3.8 GBmodel-00008-of-00015.safetensors 3.8 GBmodel-00001-of-00015.safetensors 2.5 GBmodel-00015-of-00015.safetensors 891 MBmodel-mtp.safetensors 849 MB有文件？提交种子。
提交种子