grug-v1.1-qwen-3.8-27b-mtp
same grug as ProCreations/grug-v1.1-qwen-3.8-27b, plus a draft head that guess ahead, tuned on grug own output.
Qwen3.8 ship native MTP (multi-token prediction) head. it predict token t+2 so engine can guess ahead and verify many token in one pass. speculative decoding only go fast when draft head agree with real model. grug retune that head on grug distribution.
what change
| metric | Qwen3.8 native head | after grug tuning |
|---|---|---|
| t+2 agreement with verifier | 90.04% | 95.37% |
| gold top-1 | 85.50% | 88.20% |
| eval loss | 0.3802 | 0.1649 |
disagree rate go from 1-in-10 to 1-in-21. old grug v1 mtp sit at 89.3%.
honest part: native head was already 90.0% before any training. grug not rescue broken head — grug sharpen already-good head. and gold top-1 go up too (85.5 -> 88.2), so head get genuinely better at real token, not just better at copying verifier mistake.
also honest: 95.37% is top-1 agreement with the verifier on grug spans — the quantity that decide whether a speculative draft get accepted. it is NOT a measured wall-clock speedup. real tok/s depend on your engine, batch size and draft length.
how grug train it
- backbone frozen. only the 425M-param
mtp.*module move. - teacher = grug-v1.1-qwen-3.8-27b own greedy next-token choice (self-distillation). this target acceptance directly: acceptance IS agreement with verifier.
- data = ProCreations/grug-27b-v2-corrective — agent trajectory, tool call, grug think, code, math. the distribution grug actually make in real use.
- 4.0M token seen, 1.25M supervised, 489 step, 26 minute on one RTX PRO 6000.
- verified model output unchanged — same weight, same score as v1.1. draft head only make it faster, never change answer.
full receipt in mtp_training.json.
use grug
engine that know Qwen3.8 MTP (vLLM, sglang) pick head up from config.json
(mtp_num_hidden_layers: 1) and model-mtp.safetensors.
vllm serve ProCreations/grug-v1.1-qwen-3.8-27b-mtp \
--speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}' \
--reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
plain transformers ignore mtp.* and just run the normal model — safe, only no
speedup.
GGUF with draft head: ProCreations/grug-v1.1-qwen-3.8-27b-mtp-gguf. no draft head, smaller: ProCreations/grug-v1.1-qwen-3.8-27b-gguf.
benchmark score, effort table and chart live on the main card — this variant score identical, it is the same verified weight.
apache-2.0.