1. Home
2. Guides
3. Open weights

Qwen models explained: Qwen3.8 Max, Flash and 27B, open weights vs API, with release dates

Alibaba put its Max-class Qwen out as open weights for the first time, and skipped open weights for Qwen3.7 entirely.

By Think Facility · September 23, 2026 · Qwen Alibaba open weights models

The newest Qwen models are the Qwen3.8 family, which Alibaba's Qwen team released as open weights between August 12 and August 26, 2026. The biggest one, a 2.4-trillion-parameter model, is the same model behind the paid Qwen3.8-Max API.

That was the first surprise. Qwen's own model card says "For the first time, Qwen3.8 brings a Qwen-Max-class model to open release." The second was a gap: there are no Qwen3.7 weights at all. Everything here comes from the Qwen team's GitHub repos, its Hugging Face model cards and Alibaba Cloud's price list, all read on September 23, 2026.

Every recent Qwen release, by date

Qwen ships two kinds of model. Open-weight models are files anyone can download and run, and API models are hosted by Alibaba and billed per token. Some Qwen3.8 models are both.

Open-weight dates from the News sections of the QwenLM/Qwen3.8 and Qwen3.8-Flash-Next repos. Qwen3.7's months come from the snapshot names in Alibaba's API model IDs.

If you've seen August 5 given as the Qwen3.8-27B date, that's when the Hugging Face repository was created. Qwen's own announcement of it came on August 14.

How to read a Qwen model name

Open-weight Qwen names carry their size. A plain number like 27B is a dense model with 27 billion parameters. Two numbers, like 2.4T-A95B, mean a mixture of experts: 2.4 trillion parameters in total, of which 95 billion are active for any one word. That second number is closer to what it costs to run.

API names use tiers instead. Max is the top, Plus is the middle and Flash is the cheap, fast one. Alibaba also pins versions with dated snapshots: qwen3.7-max currently points to qwen3.7-max-2026-05-20, and there's a qwen3.8-max-0902. Pin a snapshot if you don't want the model changing under you.

Qwen3.8-Max, and the 2.4T open model behind it

Qwen3.8-2.4T-A95B is the flagship. Its model card says Qwen3.8-Max, the hosted version, is "the official version based on" it, adding vision input, a non-thinking mode and a million-token window by default. The open model reads 262,144 tokens natively and stretches to 1,010,000.

You won't run this one at home. Hugging Face puts the repository at about 4.9 terabytes, and the FP8 copy (weights stored at half the precision) is still about 2.5 terabytes.

On Alibaba's API, Qwen3.8-Max costs $2 per million input tokens and $6 per million output in the international Singapore region. That's cheaper than Qwen3.7-Max, which is still listed at $2.50 and $7.50. The same page lists the open model's own ID, qwen3.8-2.4t-a95b, at the same $2 and $6.

The license is Qwen's own. Anyone can use, modify, host and sell it, but a product with more than 100 million monthly users or $20 million in monthly revenue has to show the model's name in its interface. A company running a model-as-a-service or AI work assistant business needs a separate license from Qwen once it passes $50 million in revenue over 12 months.

Qwen3.8-Flash-Next and Qwen3.8-Flash

Flash-Next is the odd one out. It's built on a new design, and the model card calls it an "experimental preview of the architecture that will underpin Qwen4". That's the only official word on Qwen4 I could find.

Its size is unusual too. The card lists "125B with 6B activated, plus 51B n-gram embedding and 4B MTP", which adds up to about 180 billion parameters on disk but only 6 billion doing work for each word. N-gram embeddings are a large lookup table keyed on short runs of tokens, and Qwen's reason for them is that they grow the model without adding much compute.

The hosted Qwen3.8-Flash is based on it and costs $0.15 in and $0.47 out per million tokens, with a million-token window. That makes it the cheapest of the three Qwen3.8 text models Alibaba hosts.

One licensing detail caught me out. Flash-Next's community license has the same name-display rule as the big model, but its model-as-a-service clause has no revenue threshold at all: any company in that business needs a separate license for commercial use. So the smaller model is, on this one point, the more restricted of the two.

Running it locally has had rough edges. We've covered Flash-Next hanging while loading in llama.cpp on AMD cards, and the unknown architecture error older llama.cpp builds give for it.

Qwen3.8-27B, the one you can run

This is the practical one for most people with a decent GPU. It's a 27-billion-parameter dense model that reads images and video as well as text, and it's the only Qwen3.8 under the plain Apache 2.0 license, so there are no user or revenue conditions attached.

It reads 262,144 tokens natively and can be stretched to a million. Thinking mode is on by default and can be switched off per request, and the card says it keeps its reasoning from earlier turns in a conversation unless you tell it not to. Alibaba also hosts it, at $0.50 in and $3 out.

Downloads show where people have landed. On September 23, 2026 Hugging Face counted about 9.8 million downloads of Qwen3.8-27B since it went up, against roughly 808,000 for Flash-Next and 75,000 for the 2.4T model. The small one is winning by a wide margin, which I'd have guessed, though not by a factor of more than a hundred over the flagship.

Where the Qwen API lives

The model cards point to Qwen Cloud as the official hosted service, and its sample code asks for an Alibaba Cloud Model Studio API key, the same service Alibaba's price list covers. Prices differ by region. Everything quoted here is the international column billed from Singapore, and the Beijing column is cheaper: Qwen3.8-Max drops to about $1.65 in and $4.95 out there.

Qwen Cloud's page for Qwen3.8-Flash says it speaks both the OpenAI and Anthropic API formats, and names Claude Code and Codex as tools it works with. By Qwen's account, an existing coding tool only needs a new base URL and key to use it (I haven't tried that myself).

What happened to Qwen3.7

Qwen3.7 only ever shipped as a hosted service. Alibaba's price list still carries qwen3.7-max, qwen3.7-plus and qwen3.7-flash, with snapshots dated from May 17 to July 15, 2026, but Qwen's Hugging Face account has nothing between Qwen3.6 and Qwen3.8. There's no Qwen3.8-Plus on the price list either, so if you need a middle tier on the API, Qwen3.7-Plus is it, at $0.40 in and $1.60 out for prompts up to 256K tokens.

Alibaba's price page gives each of these API models a free quota of one million tokens, valid for 90 days, in the Singapore region only.