When Alibaba launched Qwen4 Max Preview on November 11, 2026, it launched one model out of four. The launch slide showed the family: Max, Plus, Flash, and a 27B. Only Max is live, on the Token Plan as qwen4-max-preview. The other three are names without a release date.
That last tier, the 27B, is the one that changes how you think about the whole launch. Max, Plus and Flash are cloud models: you rent them. The 27B is built to run on your own machine. Same generation, opposite economics. This post covers what each size is for, which ones you can actually get today, the cloud-versus-local trade-off in concrete terms, and how the previous generation’s 27B lets you run something Qwen4-shaped locally right now.
The four sizes on the slide
The family breakdown comes from Alibaba’s launch slide, echoed in OrcaRouter’s analysis, which describes Qwen4 as a four-tier lineup where only Max has shipped.
| Size | What it is for | Status on 2026-11-11 |
|---|---|---|
| Max | Flagship. Complex coding, knowledge work, front-end builds, agentic runs | Live on Token Plan as qwen4-max-preview |
| Plus | Mid-tier. Everyday coding and reasoning below flagship cost | Named, no release date |
| Flash | Speed and cost tier. High-volume, latency-sensitive calls | Named, no release date |
| 27B | Open-weight, runs on your own hardware | Named, no release date |
Read that as a pricing ladder, not four separate products. Max is the ceiling. Plus is the workhorse most people will actually default to once it ships. Flash is the cheap-and-fast tier for bulk work. The 27B is the “own it forever” option.
One caution on the framing. Cellcog traces the family design back to Qwen3.8-Flash-Next, an open-weight mixture-of-experts model with 125B total parameters and roughly 6B active. That is lineage, not a spec sheet for Max: it tells you the direction is sparse MoE with long-context machinery, but it tells you nothing about Max’s window, speed or price.
Which ones you can use today
Exactly one: qwen4-max-preview, on the Token Plan from $3 a month. It is a preview of a proprietary flagship, so there are no open weights and no per-token rate. Plus, Flash and the 27B have not shipped, so there is nothing to call and nothing to download.
The Model Marketplace currently lists one Qwen4 preview id. If you see a page offering “Qwen4 Plus” or “Qwen4 Flash” today, it is ahead of Alibaba’s own catalog, which is the test that matters. Bookmark the marketplace and re-check it; the family will fill in on Alibaba’s schedule, not ours.
The 27B: why a local model changes the math
The 27B is the tier worth planning around, because it inverts the economics of the other three. Max, Plus and Flash bill per call, forever. The 27B runs on your GPU, so after the hardware cost your marginal price per token is your electricity.
You cannot run the Qwen4 27B yet. But its predecessor is live and open-weight, and it tells you what to expect. The Qwen3.8 27B is the open model behind the seven-hundred-tokens-per-second-on-an-RTX-5090 figure that circulated during launch week. That speed belongs to the small open model running locally, not to the cloud Max tier. It is the single most misattributed number of the launch.
The 3.8 27B is callable today on the free models lane as qwen3.8-27b:free, billing zero per token inside the daily free allowance. It is the closest thing to “Qwen4 27B locally” that exists right now, and a reasonable way to size up whether a local Qwen4 27B will fit your workload when it ships.
Cloud or local: how to actually choose
This is not a which-is-better question. It is a which-constraint-binds-you question.
Choose cloud (Max, later Plus and Flash) when capability is the binding constraint. If you need the strongest reasoning for hard coding tasks, or you want to offload a long-context job you cannot fit on your GPU, the cloud flagship wins. You pay per call, you scale to zero, and you get whatever Alibaba improves next. The cost is recurring and, during a preview, unpredictable.
Choose local (the 27B) when cost at volume or privacy is the binding constraint. If you run thousands of calls a day on tasks a 27B handles, or your code and data cannot leave your machine, local wins. You pay for hardware once and tokens are effectively free. The cost is capability: a 27B is not a flagship, and you are responsible for serving it.
| Decision | Cloud Max | Local 27B |
|---|---|---|
| Hardest reasoning tasks | Yes | No |
| Bulk high-volume calls | Expensive | Cheapest at scale |
| Code that must not leave your machine | No | Yes |
| Upfront cost | None | GPU required |
| Ongoing cost | Per call / credits | Electricity |
| Available today | Yes (qwen4-max-preview) |
Not yet; 3.8 27B is |
A concrete framing. A coding agent sending half a million tokens a day on a cloud flagship gets expensive fast. The same volume on a local 27B costs electricity. But if the 27B is not smart enough for your hardest refactor, no amount of cheap tokens fixes that. Match the tier to the task, not the task to the tier.
Most real setups end up hybrid. The strongest model in the cloud for the hard calls, a local or cheap tier for the high-volume easy ones. When Qwen4 Plus and Flash ship, that ladder gets easier to build on Alibaba’s own models.
What to do while the family fills in
If you need a flagship now, Max on the Token Plan is live. If you want to prep for the local 27B, there are two useful moves.
First, run the Qwen3.8 27B free on One AI API Key to learn what a Qwen 27B can and cannot do on your tasks. It is the same size class the Qwen4 27B will occupy, so it is a real capability preview.
Second, if you are building the local path, our estimate monthly LLM API cost guide and cost calculator show the token math you are trying to beat by going local, and choosing a model for coding agents walks the cloud-versus-local trade-offs across the tiers we serve today.
Run a Qwen 27B free while you wait for the Qwen4 family
qwen3.8-27b:free bills zero per token in the free lane, 100 requests a day, no card. Same size class as the coming Qwen4 27B, callable today.
Mistakes to avoid
- Assuming all four sizes are available. Only Max has shipped. Plus, Flash and the 27B are named, not released. Verify against Alibaba’s live marketplace before planning around any of them.
- Quoting 700 tok/s as a Qwen4 Max figure. That number is the open-weight Qwen3.8 27B running locally on an RTX 5090. It says nothing about the cloud Max preview’s speed.
- Confusing the local 27B with the cloud flagship. A 27B is not a Max. It is cheaper to run and less capable. Pick by task, not by brand.
- Going local to save money on hard tasks. Local is cheapest at volume, not at capability. If the 27B cannot do the task, free tokens are worthless.
- Building a whole workflow on one preview id. The family and the ids will change during the preview. Pin the id, log your calls, and re-run evals after announcements.
FAQ
What sizes does the Qwen4 family have?
Four, per the launch slide: Max, Plus, Flash and a 27B. Only Max has shipped, as qwen4-max-preview on the Token Plan. The other three are named without a release date.
Can I run Qwen4 27B locally?
Not yet. The Qwen4 27B is named but unreleased, with no weights published. Its predecessor, the open-weight Qwen3.8 27B, is available now and is the model behind the 700 tokens-per-second-on-an-RTX-5090 report. It is a good preview of what a local Qwen4 27B will do.
Which Qwen4 model should I use?
Max today, if you need the strongest capability and accept Token Plan pricing. Wait for Plus or Flash if your workload is everyday coding or high-volume and you can wait. Plan for the 27B if cost at scale or data privacy is your binding constraint. There is no single right answer; it depends on which constraint binds you.
Is the 700 tokens per second figure Qwen4 Max?
No. That figure belongs to the open-weight Qwen3.8 27B running locally on an RTX 5090. The cloud Max preview has no published speed figure, and launch-day reports were far slower under load.
How much does the Qwen4 27B cost?
No price is published because it has not shipped. As an open-weight model it is expected to be free to download and run on your own hardware, with the cost being the GPU and electricity. The Qwen3.8 27B is callable today at zero per token on the free lane.
The bottom line
Qwen4 is a four-size family, and Alibaba shipped only the top: Max, live on the Token Plan. Plus and Flash are the cloud tiers still to come, and the 27B is the open-weight model you will eventually run yourself, the one that flips the economics from renting to owning. Only Max is callable today, so if you need a flagship now, that is the one. If you are planning the local path, run the Qwen3.8 27B free on One AI API Key to learn what a Qwen 27B can do on your tasks before you commit hardware. And do not quote 700 tokens per second at the cloud Max; that number lives on a local GPU, not Alibaba’s servers.