grug think fast. now grug PREDICT own thought before think it. free speedup.
this is
grug-35b-v2 (v2.1) with a
trained MTP head for multi-token prediction / speculative decoding. base
Qwen3.6-27B ship MTP head; ordinary grug-35b-v2 drop it. grug graft head back AND
retrain it on grug data so draft match grug's terse output style.
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-35b-v2 - same tensors bit-for-bit).
1vllm serve ProCreations/grug-35b-mtp --max-model-len 32768 \
2 --reasoning-parser deepseek_r1 \
3 --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":1}'
gguf with MTP tensors:
grug-35b-mtp-gguf
(needs llama.cpp build with qwen3_5 MTP speculative support).
no MTP wanted? use plain
grug-35b-v2 -
same brain, smaller checkpoint. grug made by ProCreations.
