IMPLICATOR
Alibaba released Qwen3.8-27B with its maximum xhigh reasoning effort enabled by default. A request to draw a pelican riding a bicycle as an SVG used 22,276 reasoning tokens over 21 minutes. Early outside tests that left reasoning_effort unset measured the model at its slowest and most expensive setting.
What Changed
- Alibaba's Qwen3.8-27B shipped on August 14 with
reasoning_effortset toxhigh, the most expensive of its three levels, enabled by default. - A request to draw a pelican riding a bicycle as an SVG used 22,276 reasoning tokens over 21 minutes on that default. With reasoning disabled, the same request produced 3,715 output tokens in 137 seconds.
- Selecting
mediuminjects no instruction at all and returns no error, which leaves users on the maximum setting without knowing it. - One tester's coding task ran 11 hours on the default before finishing better than any self-hosted alternative he had tried. GPT 5.5 handled a similar job in about 20 minutes.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
A costly default
The August 14 release has 27.78 billion parameters and a 262,144-token native context window. Its Apache 2.0 license and a third-party four-bit file of about 17GB make it practical for high-end local machines.
The chat template offers xhigh, medium and low. Extra high injects instructions to check assumptions and alternatives. Low asks for brief, focused thought. Medium adds no instruction, making it the neutral point, and preserve_thinking is also on by default.
“If you benchmarked Qwen3.8 this week and left reasoning_effort unset, you measured its most expensive setting and reported it as the default,” Tom Turney posted. Setting medium, he added, is “a silent no-op, no error.” Selecting medium changes nothing and returns no error, leaving users on the maximum setting without knowing it.
What the setting costs
Simon Willison has used the pelican-on-a-bicycle prompt as what he calls “my own stupid benchmark” for close to two years. He ran a 17GB Q4_K_M build on a 128GB M5 Max MacBook Pro and an Nvidia DGX Spark. His prompt consumed 22,276 reasoning tokens and 21 minutes before producing 3,223 output tokens. With reasoning disabled, the same request produced 3,715 output tokens in 137 seconds. He called it the best pelican SVG he had produced from a model running on a local machine.
A much simpler request for a circle produced an animated composition with guide rings and tick marks after several minutes, rather than the requested plain shape. On a coding task, the default used 17,576 reasoning tokens against Muse Glimmer 30B’s 1,021. Both applications worked.
The longest elapsed-time example came from a pseudonymous commenter posting as SwellJoe, who ran an eight-bit conversion on dual Radeon V620 GPUs with no tuning. His coding task, already run against several other small models, was submitted as a pull request to one of his own repositories. The model completed it better than any self-hosted alternative he had tried, but took 11 hours. GPT 5.5 handled a similar job in about 20 minutes.
Get Implicator.ai in your inbox
Strategic AI news from San Francisco. No hype, no "AI will change everything" throat clearing. Just what moved, who won, and why it matters. Daily at 6am PST.
No spam. Unsubscribe anytime.
The ability behind the delay
The added computation sometimes paid for itself. The model placed bounding boxes closely around pelicans in a photograph and built a working labeling tool from one prompt. With reasoning disabled, the same tool request nearly worked but positioned the boxes incorrectly. In a separate test with reasoning enabled, the model drove a coding-agent loop through a real repository, answered an authentication question, then wrote and tested a Python transcript converter.
Alibaba’s own launch evaluation put Terminal-Bench 2.1 at 73.0 on August 14, up from 63.4 for Qwen3.6-27B, and DeepSWE 1.1 at 42.2, up from 13.3. No independent reproduction of those scores had appeared by launch day. Alibaba has not published the training corpus, training-token count, knowledge cutoff, post-training recipe or a safety evaluation.
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
A disputed diagnosis
On the model’s Hugging Face discussion page, a user posting as LuffyTheFox called the behavior “a structural defect in the temporal processing layers.” The user cited scale corrections of roughly 0.48 to 0.65 across eight tensors and claimed token use was about five times higher than it should be.
Other participants rejected that conclusion. Some said 137 reasoning tokens for a car-wash question was negligible. Others blamed the harness or chat template. No independent confirmation of the tensor diagnosis has appeared in that discussion.
Hardware also sets a hard floor. A pseudonymous commenter using the name madduci tried a 65,000-token context on a laptop with 32GB of RAM and no dedicated GPU: “it wasn't even starting thinking.” Another pseudonymous commenter, pdyc, reported six to eight tokens per second on the same class of machine.
Willison’s recommendation was more immediate: “ignore that default. Run Qwen 3.8 27B on low or even no reasoning levels at first.”
Frequently Asked Questions
What is Qwen3.8-27B's default reasoning setting?
xhigh, the most expensive of its three reasoning levels. The chat template routes to it whenever reasoning_effort is left unset, and preserve_thinking is also on by default.
How much does that default actually cost?
On a pelican-on-a-bicycle SVG prompt it consumed 22,276 reasoning tokens over 21 minutes before producing 3,223 output tokens. The same request with reasoning disabled produced 3,715 output tokens in 137 seconds. On one coding task run with no tuning, the default took 11 hours against roughly 20 minutes for GPT 5.5.
Does setting reasoning effort to medium fix it?
No. Medium adds no instruction to the system prompt at all, which makes it the neutral point rather than a middle gear. Tom Turney called it "a silent no-op, no error."
Is the model good despite the default?
The extra computation sometimes pays for itself. The model placed bounding boxes closely around pelicans in a photograph, built a working labeling tool from a single prompt, and drove a coding-agent loop through a real repository. With reasoning disabled, the same tool request nearly worked but positioned the boxes incorrectly.
Have the benchmark scores been independently verified?
No. Alibaba's own launch evaluation put Terminal-Bench 2.1 at 73.0, up from 63.4 for Qwen3.6-27B, and DeepSWE 1.1 at 42.2, up from 13.3. No independent reproduction of those scores had appeared by launch day, and Alibaba has not published the training corpus, training-token count, knowledge cutoff, post-training recipe or a safety evaluation.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
