Alibaba's small, open-weight member of the Qwen3.8 line — a dense, roughly 27-billion-parameter multimodal model that reads images and video, runs on a single GPU, and ships under Apache 2.0. The FP8 build fits in about 28GB of VRAM.
Last verified · 2026-08-14 · by Moe Ameen
Qwen3.8-27B is the small, self-hostable member of Alibaba's Qwen3.8 family, released as open weights in August 2026 — roughly ten days after the August 3 announcement of the flagship Qwen3.8-Max. Where the Max tier is a 2.4-trillion-parameter data-center model, this one is deliberately the opposite: a dense model of about 27 to 28 billion parameters that a single person can actually run. The whole model sits in one GPU's memory rather than being sharded across a rack, which is the entire point of the release.
The surprise in the release was that it is multimodal. Alongside text, Qwen3.8-27B ships with a vision encoder, so it accepts images and video in addition to prompts — you can hand it a screenshot, a photo, or a clip and have it describe, analyze, or reason over what it sees. Architecturally it uses a hybrid attention design (linear-attention "Gated DeltaNet" layers interleaved with standard gated-attention layers, the same lineage as the larger Qwen models) and carries a 262,144-token native context that can be extended toward roughly one million with scaling techniques. It also exposes "flexible thinking" — you can turn its step-by-step reasoning on or off and tune its depth per request — and was trained with multi-token prediction for faster decoding.
Two practical facts make this the Qwen3.8 build most creators will actually touch. First, the license: unlike the flagship Qwen3.8-Max weights, which ship under a custom license with signaled revenue-sharing for large commercial users, Qwen3.8-27B is Apache 2.0 — commercial use, modification, and redistribution are all permitted. Second, the hardware. The base BF16 weights need roughly 56GB of VRAM, but the official FP8 build (fine-grained FP8 quantization at block size 128, with quality reported as near-identical to the original) drops that to about 28GB, so it fits on a single 48GB card, and community 4-bit builds bring it into 24GB-class GPU range. Alibaba positions the model for coding, professional and research work, multimodal understanding, and long-horizon agentic tasks. If a specific benchmark number, the exact release date, or a hardware figure matters to your build, confirm it on the official Qwen model card before you commit — this is a fast-moving line and vendor benchmark scores are not yet independently verified.
The distinctive thing about Qwen3.8-27B is that it can see — its vision encoder reads your images and watches your clips — and that is exactly the boundary worth naming: it can look at media, but it cannot make any. Hand it a photo and it will describe the lighting; hand it a rough cut and it will summarize the beats. What it will never do is render the reframed vertical video, design the branded carousel, or produce the quote card those inputs deserve. It is an analyst and a drafter that happens to run on one GPU in your own office. [Kompozy](/) is the production layer that turns everything it sees and drafts into finished, on-brand assets. Point your self-hosted Qwen endpoint at Kompozy through bring-your-own-key on the Founding tier, and its analysis and copy become [Persona Shorts](/glossary/persona-shorts) and HeyGen avatar video, brand-exact Carousels and Persona Tweets rendered through [HyperFrames](/glossary/hyperframes), Photo Posts, Quote Graphics, a blog article, and an email newsletter — the media a vision model can only look at.
The pairing is efficient because the two halves genuinely do not overlap. Run the model where its scale earns its keep — private, unmetered drafting and letting it read your own footage and screenshots to surface angles — then let Kompozy cast that raw material into your voice through the [Persona Brief](/glossary/persona-brief) and banned-word filters, reframe each output to 9:16, 1:1, and 16:9, and schedule and publish the whole set across the eight social platforms plus blog and email with a per-post review pipeline and [Autopilot](/glossary/autopilot). Because it is Apache 2.0 and fits on a single card, a solo creator or small team can adopt it for cost and privacy without a data-center budget — and keep the entire generation stack in-house right up to the render-and-ship boundary Kompozy owns. Qwen3.8-27B reads and drafts on your own hardware; Kompozy makes it look like your brand and ships it everywhere. (One note: Kompozy's own copy generation runs on managed Claude and OpenAI, so a self-hosted Qwen is a private drafting-and-vision brain feeding the pipeline, not a swap for it.)
Qwen3.8-27B is the small, open-weight member of Alibaba's Qwen3.8 line, released in August 2026. It is a dense model of roughly 27 to 28 billion parameters — the flagship Qwen3.8-Max is a 2.4-trillion-parameter model by contrast — and it is multimodal, with a vision encoder that reads images and video alongside text. It carries a 262K-token native context and is built to run on a single GPU.
Yes. The weights are downloadable from Hugging Face and ModelScope under Apache 2.0, which permits commercial use, modification, and redistribution. That is notably more permissive than the flagship Qwen3.8-Max weights, which ship under a custom license with signaled revenue-sharing for large commercial users. Confirm the license on the official model card for your specific build.
The base BF16 weights need roughly 56GB of VRAM, so an 80GB card. The official FP8 build (block-size-128 quantization with near-identical quality) drops that to about 28GB, fitting a single 48GB GPU, and community 4-bit builds run in the 24GB range of a high-end consumer card. KV cache for long context adds to those figures, so budget headroom.
Yes — this build includes a vision encoder and accepts images and video in addition to text, which was a surprise in the release. It can describe a photo, read a screenshot, or reason over a clip. It is an input capability: the model understands media you give it, but it does not generate video, images, or audio of its own.
No. It drafts and reasons over text and reads images and video, but it produces no finished video, images, or designs, enforces no brand voice, and publishes to no platform. To turn its drafts and analysis into on-brand posts across platforms, pair it with a content engine like Kompozy — which can call your own self-hosted Qwen key on the Founding tier — that renders the media and handles scheduling and publishing.