grug-v1.1-qwen-3.8-27b

grug think small, answer big.

normal model think like this:

Okay, so the user wants me to carefully consider the best approach here. Let me think about this step by step. First, I should consider what data structure...

grug think like this:

Sort numbers once; adjacent differences sufficient. Any diff < threshold -> True; else False. Edge len < 2 -> False.

same reasoning. same steps. same answer quality. grug throw grammar padding in fire, keep brain meat. answer come out normal english — only inside voice is grug.

new rock under grug: Qwen3.8-27B. old grug sit on Qwen3.6.

grug hold quality

grug spend fewer token to think

base model burn 559 token thinking about HumanEval. grug burn 79.5. same kind of answer. on agent step base burn 108.5, grug burn 20.

grug pick right tool

this is the big one. base model can call a tool — it call a valid tool 98.5% of time. but base pick the right tool only 23.5% of time. grug pick right tool 97.1%.

numbers, all of them

medium reasoning effort, full benchmark sets (HumanEval 164, MBPP 100, GSM8K 200, MATH-500 150, agentic 68, recovery 80, repetition 43). same harness, same settings, every column.

mean reasoning token per answer:

grug across reasoning effort

Qwen3.8 template take reasoning_effort. same weight, three setting. grug measure all three on this exact released model, full sets:

more effort is NOT more better. medium win 5 of 7 probe. xhigh actively hurt tool pick — 76.5 vs 97.1, drop 20 point, because xhigh instruction tell model "think carefully, consider alternative", and that push agent toward writing essay instead of calling tool. base model also get worse at xhigh (HumanEval 98.2 -> 92.7). so grug say: use medium.

low is real budget option though: 464 think token total instead of 680 (-32%), and still 85.0 MBPP / 93.0 repetition. pay ~7 point HumanEval for it.

(old grug v1 cannot do this at all — Qwen3.6 template ignore reasoning_effort, all three setting render byte-identical prompt. dial is new.)

where grug lose

grug not hide bruise.

• GSM8K 92.5 — base 95.5, old grug 96.5. grug lose 4 point on grade-school word problem. real loss, not noise.
• recovery right-tool 82.5 — old grug 90.0. when tool fail and grug must pick next action, old grug pick better.
• MBPP think 301 token — old grug spend 34.6. grug think much longer here for +4 point. if you want cheapest possible token on MBPP, old grug cheaper.
• base still better at raw code and math. grug trade some point for 7-30x fewer think token. if token free and you want max score, use base.

why called v1.1 and not v2

because gain over v1 is honest-small. new base rock, big win on tool choice and repetition and MATH-500, but GSM8K step back. that is a point-one, not a two. grug not put big number on small step.

how grug get made

1. base Qwen/Qwen3.8-27B
2. SFT on 1M-row grug corpus (grug think, normal answer)
3. corrective LoRA (rank 32) on top, then applied at 0.5 strength

step 3 matter more than it sound. at full strength the adapter overshoot — it push so hard toward tool behaviour that it break code:

less adapter = better code AND better tool pick, both at once. grug learn: more push not always more better.

three bug grug step on, so you not have to

first build of this model score fine on code and then collapse on tool call — 20.6% valid call. three rot in training data, none in model:

1. every agent story end with "all done" speech. all 19,129 agent row finish with assistant turn that make no tool call, just summary sentence. model learn "deep in trajectory = time to stop". so when judge ask for next tool call deep in hunt, model write essay. fix: re-cut same story so supervised turn IS a tool call.
2. think was english, not grug. function-word ratio 0.10-0.29 where v1 sit at 0.01-0.09. that why first build burn 203 token on HumanEval.
3. double think. Qwen3.8 template fill own <think> block from reasoning_content field. old data hid think inside content as <think> tag. so every row train as <think>\n\n</think> then second literal <think> — empty think, then duplicate. 59% of supervised turn carried empty think. that why model sometimes shut reasoning instantly then reason inside answer.

grug no loop

old 35b brother repeat word until cave fall down. grug train on BOTH world: think-in-history trajectories AND stripped-history variants (old think gone, exactly like real agent framework replay). then repetition stress gauntlet before release: greedy long-form, deep think-stripped agent replay, multi-turn continuation. zero loop, 100% think closed.

use grug

from transformers import AutoTokenizer, AutoModelForImageTextToText

tok = AutoTokenizer.from_pretrained("ProCreations/grug-v1.1-qwen-3.8-27b")
model = AutoModelForImageTextToText.from_pretrained(
    "ProCreations/grug-v1.1-qwen-3.8-27b", dtype="bfloat16", device_map="auto")

messages = [{"role": "user", "content": "write a function that flattens a nested list"}]
ids = tok.apply_chat_template(messages, add_generation_prompt=True,
                              reasoning_effort="medium", return_tensors="pt")
print(tok.decode(model.generate(ids.to(model.device), max_new_tokens=512)[0]))

use reasoning_effort="medium". every number on this card is medium. grug tuned there. low and xhigh work but are not what grug measured for release.

GGUF: ProCreations/grug-v1.1-qwen-3.8-27b-gguf (with mmproj for vision).

honest small print

• tool call use XML <function=> shape, same as grug v1.
• agentic and recovery probe sample at temperature 0.6. over three repeat run the right-tool number move about 1 point (stdev 0.9). numbers here from one complete error-free run each.
• every score on this card come from one open harness, same settings for base, v1 and v1.1. no cherry-pick between runs.

apache-2.0, like the rock it stand on.