Alibaba · preview runs only
Qwen 3.8 27B and Taskade
Qwen 3.8 27B from Alibaba is not in the Taskade model menu. We ran it in preview tests, which the TSK-1 ranking does not count. Its newest preview run was on Oct 23, 2026. TSK-1, our app-building test, does not rank preview runs.
- Status
- Not in Taskade
- Context window
- 262KtokensHow much it can read at once
- Max output
- 33KtokensLongest answer it can write
- Released
- Where it runs
- 9servicesOther services that serve it
Published scores
Benchmark scores that Alibaba and test sites published for Qwen 3.8 27B. Taskade did not re-run them. Lab-reported marks a score from Alibaba itself.
| Benchmark | Score | Source |
|---|---|---|
| Terminal-Benchv2.1 · in TerminusHugging FaceLab-reported | 73 | Hugging FaceLab-reported |
| LiveCodeBenchv6Hugging FaceLab-reported | 90.3 | Hugging FaceLab-reported |
| OSWorld-VerifiedHugging FaceLab-reported | 84.3 | Hugging FaceLab-reported |
| GPQA DiamondHugging FaceLab-reported | 89.2 | Hugging FaceLab-reported |
| Humanity's Last Examwithout tools · GPT-4o judgeHugging FaceLab-reported | 30.8 | Hugging FaceLab-reported |
More published scores (25)
| Benchmark | Score | Source |
|---|---|---|
| SWE-bench ProresolvedHugging FaceLab-reported | 61.7% | Hugging FaceLab-reported |
| NL2Repoin Claude CodeHugging FaceLab-reported | 42.3 | Hugging FaceLab-reported |
| DeepSWEv1.1 · in Claude CodeHugging FaceLab-reported | 42.2 | Hugging FaceLab-reported |
| QwenSWEBenchaverage of 3 runs · in Claude Code · the lab's own test setHugging FaceLab-reported | 79 | Hugging FaceLab-reported |
| CoWorkBenchthe lab's own test setHugging FaceLab-reported | 70.7 | Hugging FaceLab-reported |
| JobBenchHugging FaceLab-reported | 33.4 | Hugging FaceLab-reported |
| Agents' Last Examfirst tryHugging FaceLab-reported | 20.4% | Hugging FaceLab-reported |
| Agents' Last ExamHugging FaceLab-reported | 42.9 | Hugging FaceLab-reported |
| IFBenchHugging FaceLab-reported | 79.5 | Hugging FaceLab-reported |
| WebArena-Verifiedin OSWorld scaffold; official WebArena-Verified graderHugging FaceLab-reported | 64.8 | Hugging FaceLab-reported |
| AndroidWorldHugging FaceLab-reported | 81.9 | Hugging FaceLab-reported |
| SWE-MMHugging FaceLab-reported | 38.6 | Hugging FaceLab-reported |
| RealWorldQAHugging FaceLab-reported | 85.9 | Hugging FaceLab-reported |
| ERQAHugging FaceLab-reported | 65.5 | Hugging FaceLab-reported |
| RecreationBenchthe lab's own test setHugging FaceLab-reported | 47.1 | Hugging FaceLab-reported |
| ClawEval-MMbest of 3 triesHugging FaceLab-reported | 57.4 | Hugging FaceLab-reported |
| ClawEval-MMaverage scoreHugging FaceLab-reported | 56.9 | Hugging FaceLab-reported |
| Vision2Webgpt-5.4-2026-03-05 judge · in Claude CodeHugging FaceLab-reported | 62.9 | Hugging FaceLab-reported |
| MathVisionwithout code interpreter · corrected annotationsHugging FaceLab-reported | 90 | Hugging FaceLab-reported |
| MathVisionwith code interpreter · corrected annotationsHugging FaceLab-reported | 94.6 | Hugging FaceLab-reported |
| BabyVisionwithout code interpreterHugging FaceLab-reported | 65.7 | Hugging FaceLab-reported |
| BabyVisionwith code interpreterHugging FaceLab-reported | 85.6 | Hugging FaceLab-reported |
| CharXiv Reasoningwithout code interpreter · corrected annotationsHugging FaceLab-reported | 83.7 | Hugging FaceLab-reported |
| CharXiv Reasoningwith code interpreter · corrected annotationsHugging FaceLab-reported | 90.2 | Hugging FaceLab-reported |
| OmniDocBenchv1.5Hugging FaceLab-reported | 91.1 | Hugging FaceLab-reported |
LMArena scores
| Benchmark | Score | Source |
|---|---|---|
| Text arena±5 · 27,779 votes · Rank 106 of 419 · LMArena | 1,436 | LMArena |
| WebDev arena±6 · 18,752 votes · Rank 34 of 152 · LMArena | 1,590 | LMArena |
Scores from the LMArena leaderboard dataset (CC BY 4.0). People vote between two answers, and the votes set the score. Taskade picked the Overall category and the default setting of each model.
Scores collected by models.dev (MIT). Each source links the page that published the score.
Versions tested
Preview runs only, not ranked
Model facts
Published specs from the model maker. A token is about three quarters of a word. Taskade can run the model with lower limits than these.
- Last updated
- Reads
- Text, Images, PDFWhat you can give it
- Writes
- TextWhat it gives back
- Reasoning
- YesThinking a problem through first
- Tool calling
- YesUsing tools, such as web search
- File attachments
- YesReading files that you attach
- Structured output
- YesFilling a set format, such as a table
- Open weights
- YesFree for anyone to download and run
- Thinking levels
- NoneLowMediumExtra highHow long it can think. Higher is slower and more careful.
Published specs from the model maker, via models.dev (MIT).
Where it runs
Services outside Taskade that serve Qwen 3.8 27B, from models.dev. Each link opens in a new tab.
What to pick in Taskade
- Leave TSK-Auto on, and Taskade picks a current model for each request.
- To pick a model yourself, choose one from the model menu.
- To compare, open the models page and read the test results of the models you can pick.
Compare Qwen
Other Alibaba models
FAQ
Can I pick Qwen 3.8 27B in Taskade?
No. Qwen 3.8 27B is not in the Taskade model menu, and the key providers in Taskade do not list it by name. TSK-Auto picks a current model for each request. The models page lists the models you can pick now.
Has Taskade tested Qwen 3.8 27B?
Only in preview runs, last on Oct 23, 2026. The TSK-1 ranking does not count them, so Qwen 3.8 27B has no TSK-1 score. The TSK-1 page shows what each run built.
Who makes Qwen 3.8 27B?
Alibaba makes Qwen 3.8 27B. Taskade runs frontier models from top AI labs, and the Qwen results page shows how Alibaba models build apps.