Qwen3.8 arrives in the open: a runnable 27B, and a text-only 2.4T flagship
Originally announced by Qwen (Hugging Face) . We link the primary source so you can read it for yourself.
When we covered the Qwen3.8-Max announcement on 3 August, the weights were not yet published and no licence had been named. We said to treat both as announced rather than available. They have now landed, and the picture is both clearer and more interesting than a single release. Qwen3.8 arrived as two very different things.
The one worth running: Qwen3.8-27B
For anyone running local AI on their own hardware, Qwen3.8-27B is the release that matters. It is a dense 27-billion-parameter model, so there is no mixture-of-experts machinery and no cluster involved: it loads and runs like any other single-card model. At a 4-bit quantisation it fits a 24GB card, which puts it within reach of a used RTX 3090 or a well-specified Mac.
Two things lift it above a routine size bump. It is natively multimodal, reading images and video as well as text, which is unusual in a model this size that is also a capable general assistant. And it carries a 262k context window that extends towards a million tokens with YaRN scaling. It is published under Apache 2.0, so there are no commercial-use conditions to weigh.
Qwen reports strong evaluation scores for it, including 90.3 on LiveCodeBench v6, 89.2 on GPQA Diamond, and 61.7 on SWE-bench Pro. As always, these are the developer’s own figures rather than independent results, and are worth treating as a starting point until third-party evaluations appear. Even discounted for that, a permissively licensed, multimodal 27B that runs on one card is a genuine step for local users.
The flagship: open, but text-only and cluster-scale
The headline number belongs to the other release. The open-weight checkpoint of Qwen3.8-Max, named Qwen3.8-2.4T-A95B, is a 2.4-trillion-parameter mixture-of-experts activating 95 billion parameters per token. That is frontier-scale, and it is genuinely open. But the detail matters, and it cuts against the local-AI grain in two ways.
First, the open checkpoint is text-only. The model card states plainly that multimodal input is not supported, so the vision that the hosted Qwen3.8-Max offers is not part of what was released. It also requires thinking mode for every interaction, which adds latency and token use. Second, it was published in BF16 at roughly 4.8TB of weights. This is multi-node cluster territory, in the same class as Kimi K3: a model to serve from organisational infrastructure, not to run at a desk. If you have that infrastructure and want to self-host a frontier-scale open model with your data in-house, it is a real option, and we have added it to the catalogue as an example of that server-class tier.
A notable licence change
There is one more thing worth flagging, because Qwen’s licensing has been a reason to prefer the family. The 27B is Apache 2.0, consistent with the rest of the line. The flagship is not: it ships under a custom Qwen3.8-Max licence. Reporting at release described commercial use as permitted but requiring a separate agreement above a large revenue threshold, in the manner of Meta’s Llama licence. We have not independently confirmed the exact terms, so anyone considering the flagship commercially should read the licence on the model card directly. It is a departure worth noting, even if it affects only deployments at real scale.
What to take from it
The useful summary is that Qwen3.8 gave local users one clear win and one model most will admire rather than run. The 27B is permissively licensed, multimodal, and runnable on a single card, and it is the one to try. The 2.4T flagship is an impressive open release, but a text-only, cluster-scale one under a custom licence, and the smaller model is where the practical value sits.