Results
Scored once, at the end, on 450 held-out problems (150 each from GSM8K, MATH-500 and MMLU-Pro) that were never used
to build or choose the edit. Thinking mode, temperature 1.0, top-p 0.95, top-k 20, up to 16,384 new tokens, 2
sampling seeds per model, compared problem by problem.
Table with columns: Task, Thinking, Qwen3.8-27B, Thinking, Quickthinker, Change, Accuracy, Qwen3.8-27B, Accuracy, Quickthinker| Task | Thinking, Qwen3.8-27B | Thinking, Quickthinker | Change | Accuracy, Qwen3.8-27B | Accuracy, Quickthinker |
|---|
| GSM8K | 528 | 164 | −69% | 0.980 | 0.977 |
| MATH-500 | 1,897 | 1,073 | −43% | 0.963 | 0.990 |
| MMLU-Pro | 3,429 | 2,110 | −38% | 0.797 | 0.817 |
| All | 1,951 | 1,116 | −42.8% | 0.913 | 0.928 |
Thinking is counted as every token up to and including </think>. A run that never closes its thinking counts as
wrong.
Read the accuracy as "the same", not "better". The +1.4 points (95% CI −0.1 to +3.0) come almost entirely from
fewer runs hitting the token cap. On problems where every run of both models finished, accuracy moved by −0.3 points
on GSM8K, +0.7 on MATH-500 and 0.0 on MMLU-Pro. On the 30 hardest validation problems with a 32k cap, the two models
scored the same within noise (0.583 vs 0.567, CI ±10 points) and Quickthinker still used 40% fewer tokens.
Excluding every problem where any run was truncated, the thinking cut is −50%, so it is not an artifact of truncation.
Quickstart
Use it exactly like Qwen3.8-27B. The weights are about 54 GB in bf16, so you need one 80 GB-class GPU or larger.
vLLM
vllm serve scar-ai/Qwen3.8-27B-Quickthinker \ --reasoning-parser qwen3 \ --max-model-len 32768
from openai import OpenAI client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")r = client.chat.completions.create( model="scar-ai/Qwen3.8-27B-Quickthinker", messages=[{"role": "user", "content": "A train leaves at 9:40 and arrives at 13:05. How long is the trip?"}], temperature=1.0, top_p=0.95, max_tokens=16384, extra_body={"top_k": 20},)print(r.choices[0].message.reasoning_content) # the (shorter) thinkingprint(r.choices[0].message.content) # the answer
import torchfrom transformers import AutoModelForImageTextToText, AutoProcessor repo = "scar-ai/Qwen3.8-27B-Quickthinker"processor = AutoProcessor.from_pretrained(repo)model = AutoModelForImageTextToText.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto") messages = [{"role": "user", "content": [{"type": "text", "text": "Is 391 prime?"}]}]inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=16384, do_sample=True, temperature=1.0, top_p=0.95, top_k=20)print(processor.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Build it from your own copy of Qwen3.8-27B
The whole edit ships separately as wsp_patch.safetensors (144 KB): the 12 new lm_head rows and 2 new
down_proj columns. apply_release.py writes the edited model from a local copy of the original. The three
affected shards are rewritten and the rest are symlinked. The output is bit-identical to this repository.
hf download Qwen/Qwen3.8-27B --local-dir Qwen3.8-27Bhf download scar-ai/Qwen3.8-27B-Quickthinker wsp_patch.safetensors apply_release.py --local-dir .python apply_release.py Qwen3.8-27B Qwen3.8-27B-Quickthinker wsp_patch.safetensors
How it was made
Weight-space programming
Training changes a model by showing it examples and letting an optimizer adjust billions of weights.
Weight-space programming treats the network more like a program: find where a behavior is represented, measure
it as a direction in activation space, and write the change into specific weights in closed form. Every edit is
small, has a formula, and can be inspected and reverted.
The work was run end to end by miniml, our in-house autonomous research harness, on one AMD MI300X in about
12.5 hours (about 22 GPU-hours). miniml framed the metric, built disjoint data splits, measured seed noise, ran 21
experiments and about 30 full vLLM evaluations, confirmed wins with extra seeds, and pre-registered its selection
rule before opening the test set. In weight-space mode its guardrails block optimizers and backpropagation, cap the
forward passes spent on analysis, and lock the held-out set. A separate reviewer agent audited the result twice.
The three edits, what did not work, and the limitations are covered in the
write-up.
License
Apache-2.0, the same as Qwen/Qwen3.8-27B. All credit for the base model
goes to the Qwen team.
Citation
@misc{reflexlabs2026quickthinker, title = {Qwen3.8-27B-Quickthinker: 43\% less thinking from a 0.00027\% weight edit}, author = {{Reflex Labs}}, year = {2026}, url = {https://huggingface.co/scar-ai/Qwen3.8-27B-Quickthinker}}