Qwen3.8: What's New, How to Run It, and Alternatives (2026)
Alibaba's Qwen team released Qwen3.8-27B on August 14, 2026, two days after the much larger Qwen3.8-2.4T-A95B. The release note is one sentence long and worth reading twice.
"For the first time, Qwen3.8 brings a Qwen-Max-class model to open release."
That is a claim about tier, not about size. It says the open weights now sit where the hosted flagship used to sit.
Whether that holds for your workload is a separate question, and the published evidence only partly answers it. The rest of this piece is about which parts.
This piece covers what shipped and where the two official documents disagree. It then looks at what the benchmark table actually measures. Finally it covers how to run the model locally, if you want the analysis to stay on your own machine.
What Qwen3.8 actually is
The official README describes the family as built "on the architectural foundation of Qwen3.5," delivering gains "across coding, professional work, research, and long-horizon agentic tasks."
The sentence that matters for anyone doing multi-step work is the next one. Qwen3.8 "is designed to carry complex, multi-step tasks through to completion with greater reliability."
The 27B model is the one most people will run locally. Its model card gives the hard numbers: 27B parameters, 64 layers, hidden dimension 5120, and a hybrid layout of Gated DeltaNet and Gated Attention blocks.
Context length is 262,144 tokens natively and extensible up to 1,000,000. The license recorded on the model card is apache-2.0.
The two release entries are worth keeping straight. The 2.4T-A95B mixture-of-experts model landed on 2026-08-12, and the 27B dense model followed on 2026-08-14. Only the second one is realistic to self-host.
| Fact | Qwen3.8-27B |
|---|---|
| Released | 2026-08-14 |
| Parameters | 27B dense |
| Layers | 64 |
| Context | 262,144 native, extensible to 1,000,000 |
| License | apache-2.0 |
| Inputs | Text, images, video |
Where the two official documents disagree
This is the part almost nobody reports, and it changes what you can safely claim.
The GitHub README attributes the multimodal architecture to Qwen3.5, the generation Qwen3.8 is built on. Read only that page and you would conclude the 27B model is text-only.
The model card says otherwise, and it is specific. Qwen3.8-27B is "a native vision-language model that understands images and videos, with flexible thinking control."
When two official sources conflict, the more specific one wins. The model card is describing this exact checkpoint, so the multimodal capability is real. It is worth knowing the discrepancy exists, because a quick skim of the repo will tell you the opposite.
What the benchmark table measures, and what it leaves out
The model card publishes a comparison against Qwen3.6-27B, Qwen3.7-Plus, Muse Glimmer-30B, and Opus4.6 Max. Some of the agentic numbers are large jumps.
| Benchmark | Qwen3.8-27B | Qwen3.6-27B |
|---|---|---|
| OSWorld-Verified (computer use) | 84.3 | 63.9 |
| WebArena-Verified (browser use) | 64.8 | 48.8 |
| AndroidWorld (mobile use) | 81.9 | 70.3 |
| SWE-bench Pro (coding) | 61.7 | — |
| MathVision (with code interpreter) | 94.6 | 90.3 |
| Vision2Web (visual web dev) | 62.9 | 45.0 |
Now the honest reading. Computer use, browser use, mobile use, coding, visual maths, and web development are what this table covers.
There is no benchmark here for reading a spreadsheet, reconciling two exports, or producing a report from tabular data. Strong OSWorld numbers tell you the model can drive a desktop. They do not tell you it will get your revenue reconciliation right.
One detail worth flagging for readers who follow open releases. Muse Glimmer-30B appears in this table as a comparison point, and its own launch table four days earlier compared against Qwen3.6-27B. Our piece on Muse Glimmer noted that its peer set was already a generation behind at launch. This table is the other half of that story.
How to run it
Weights are on the Hugging Face Hub and on ModelScope, per the README's news entry. The 27B dense shape is the practical choice for a single machine.
Plan your memory around the parameter count rather than the context ceiling. A 262,144-token window is available, but filling it costs memory that most local setups do not have spare.
Start with a short context and a real file. Load one export, ask one question you already know the answer to, and check the answer before trusting anything longer.
That verification step is not optional for tabular work. A model can describe a spreadsheet fluently and still misread which column holds the totals. Fluent wrong answers are harder to catch than obvious ones.
If you need the million-token window, treat it as a separate exercise with its own hardware budget. The two configurations behave very differently in practice, and benchmarking one tells you little about the other.
Change the default reasoning effort first
Here is the setting that will shape your first impression, and it is visible in the model card's own chat template.
The template resolves reasoning_effort to xhigh by default, with medium and low as the other accepted values. Thinking can also be turned off.
An xhigh default means the model will deliberate at length on questions that do not need it. For a one-line lookup against a small CSV, that reads as the model being slow rather than careful.
So the first thing to tune is not the prompt. Drop the effort to medium or low for routine questions, and reserve xhigh for the multi-step work the release note is actually about.
Where this fits in a data workflow
An open-weight model you host yourself solves one specific problem, which is that the data never leaves your machine. For regulated or sensitive files that is the whole argument, and no amount of hosted convenience replaces it.
Everything else about the choice is a trade. You are taking on hardware, updates, and quantisation decisions in exchange for that one guarantee.
What it does not solve is everything around the model. Column profiling, joining two exports, catching a renamed field, rendering a chart, and producing a document are separate pieces of work.
That gap is why protocol and tooling layers exist. Our explainer on what MCP is covers the standard that connects models to tools. The data connectors page shows what a file-first pipeline expects as input.
Alternatives
If the requirement is local and open, Qwen3.8-27B and Muse Glimmer-30B are the two current candidates in the same size class. Both publish weights and both run on a single machine.
If the requirement is a finished artifact from a file rather than a model endpoint, the shape of the tool is different. Powerdrill Bloom takes the spreadsheet, profiles the columns, and returns the chart, the report, or the deck.
There is no pricing comparison to make on the Qwen side. No official Qwen pricing page was reachable for this model, so this article quotes no figures.
If your actual task is a file that needs to become a report, try Powerdrill Bloom on it directly. See also our guide to turning a subscription export into an MRR and ARR report and the CSV AI assistant page.
Conclusion
Qwen3.8-27B is a 27B dense model with a 262,144-token native context, apache-2.0 weights, and documented image and video input. The agentic jumps over Qwen3.6-27B are substantial.
Two caveats travel with it. The repo and the model card disagree on multimodality, and the published benchmarks cover desktop, browser, coding, and visual tasks rather than tabular work.
Set reasoning_effort before you judge it. The default is xhigh, and that default is doing more thinking than most questions require.
Frequently asked questions
When was Qwen3.8 released?
The README's news entries record Qwen3.8-27B on 2026-08-14 and Qwen3.8-2.4T-A95B on 2026-08-12, both on the Hugging Face Hub and ModelScope.
Is Qwen3.8-27B multimodal?
Yes, according to its model card, which describes it as a native vision-language model that understands images and videos. Note that the GitHub README attributes the multimodal architecture to Qwen3.5.
What context length does it support?
The model card states 262,144 tokens natively, extensible up to 1,000,000. Running near the upper figure requires memory well beyond a typical local setup.
Why does it feel slow on simple questions?
The chat template defaults reasoning_effort to xhigh. Lower it to medium or low for routine lookups and keep xhigh for genuinely multi-step tasks.
Do the benchmarks say anything about spreadsheet work?
No. The published table covers computer use, browser use, mobile use, coding, visual maths, and web development, with no benchmark for tabular analysis or report generation.