Qwen3.8-27B: The Dense 27B in Alibaba's Qwen3.8 Line

One number decides whether you can run Qwen3.8-27B yourself: 55.6 GB. That is the BF16 weight total in the vendor's own repository — eighteen safetensors shards, 27.8 billion parameters once the vision tower is counted — and it is the reason the model's practical home is a 24 GB graphics card running a compressed build rather than a workstation running it as shipped. Everything else about Qwen3.8-27B follows from that shape. It is a dense model, not a mixture of experts: every token you generate lights up all 27 billion parameters, which is the opposite trade from the sparse giants it shares a generation with. And it is the one member of this round with a real neighbourhood around its name — readers search it on its own, and they search the qualifiers around it — so this page is the plain answer to what the thing is, what it is for, and what you are choosing before you pick a size or a format.

The precise identification, because three names in this family are easy to confuse. Qwen3.8-27B is the dense, self-hostable member of Alibaba's Qwen3.8 generation. Its siblings are different products: Qwen3.8-2.4T-A95B is the 2.4-trillion-parameter sparse mixture-of-experts whose weights also shipped in August 2026, and Qwen3.8-Max is the hosted API built on that same architecture, which has no downloadable weights at all. The generation was announced on 3 August 2026, and the 27B's weights landed on Hugging Face and ModelScope on 14 August 2026 under Apache 2.0 — a date worth stating precisely because the launch is only what dates this model, not an event this page is reporting.

What it is, in Alibaba's own words

The model card describes Qwen3.8-27B as a "compact, deployment-friendly dense model": a native vision-language model that understands images and videos, with flexible thinking control, "designed to carry complex, multi-step tasks through to completion with greater reliability." That is the vendor's positioning, attributed, and the card backs it with five claimed improvements — core capability gains across coding, professional work, research and long-horizon agentic tasks; stronger autonomous planning and handling of environment feedback; broader support for popular agent harnesses and development tools; thinking mode that is on by default but can be disabled per request with depth tuned through reasoning_effort; and native image and video understanding that the card says spans STEM diagrams and documents through to hour-scale video.

The architecture behind that claim is a hybrid. Qwen3.8-27B is laid out as sixteen repeating blocks, each containing three Gated DeltaNet linear-attention layers and one full Gated Attention layer — 64 layers total, hidden dimension 5,120, 48 linear-attention heads for values against 16 for queries. In plain terms: only one layer in four keeps a conventional attention cache. That is why a 27B model can carry a 262,144-token native context without the memory bill a conventional transformer of this size would demand, and it is also why its local footprint is friendlier than its parameter count suggests. Context extends to 1,000,000 tokens with YaRN scaling, which the card warns can degrade performance on shorter inputs if you leave it switched on statically. Thinking mode is on by default; the card's recommended sampling is temperature 1.0 with top_p 0.95 when thinking, 0.7 with top_p 0.80 when not, and it suggests allocating output generously — up to 262,144 reasoning tokens and 131,072 for the final answer inside a long window.

Now the honest general part, which is craft knowledge rather than a measurement of this model. A dense 27B in 2026 sits in a specific place. It is large enough to be genuinely useful and small enough to fit on hardware a person or a small team owns, which makes it the right shape for private inference, for batch work where per-token billing is the enemy, and for agent loops that make many cheap calls rather than one expensive one. It is not the right shape for the hardest reasoning you have, for the longest single documents, or for anything where you would rather rent state-of-the-art than own good-enough — a dense 27B loses to a well-served frontier model on the top end, and it loses to a small MoE on cost per token at the bottom. Density is also a latency fact: with every parameter active on every token, a dense model is generally slower per token than a sparse model of similar quality, and Qwen3.8-27B is a reasoning model on top of that, so it spends output tokens on thinking before it answers. None of this is a score for this particular model. It is the shape of the decision you are making.

Size and fit, in summary

The summary, with the detailed arithmetic left where it belongs. Full BF16 weights are 55.6 GB, which is an 80 GB accelerators conversation. The hybrid attention layout is what makes the compressed builds unusually tractable: a 4-bit quantisation lands in the mid-teens of gigabytes and runs comfortably on a 24 GB card at ordinary context lengths, a 16 GB card is realistic only if you accept a lower quantisation or short context, and 8 GB is out. Because only 16 of the 64 layers hold a full attention cache, long context costs less than it would on a conventional 27B — but it is not free, and the cache is what pushes a comfortable 24 GB setup toward its ceiling as you raise the window. We keep a separate page with the per-quantisation VRAM ladder, the context-length curve and the Apple-silicon unified-memory figures; if you are buying hardware, read that one instead of this one.

The formats, as a choice rather than a how-to

Four routes exist and they answer four different questions. GGUF is the compatibility choice — the llama.cpp lineage, the widest hardware support, the format that will run on whatever you already own, and it is what most people searching Qwen3.8-27B are actually looking for: pick it when you want the weights to work on the machine in front of you and you are willing to choose a quantisation tier. MLX is the Apple-silicon choice, and it is a real one — this model runs on a Mac through MLX, which for a 27B vision-language model is not something to assume. On the serving side, NVFP4 and FP8 are the formats you deploy rather than run interactively: FP8 keeps the model close to full precision at roughly half the weight footprint for a single large GPU, and NVFP4 is the aggressive Blackwell-era build that measures around 25 GB of VRAM with the vision tower kept at full precision. Unsloth is the fine-tuning route — the GGUF pack with a training path attached, for when you intend to change the weights rather than just run them. Each of those has its own page under this model, and each answers a different question than this one does.

What the licence actually permits

Qwen3.8-27B ships under Apache 2.0. The LICENSE file is in the repository, and the licence is the permissive standard rather than a custom agreement: commercial use without a user cap or a royalty, modification, redistribution, and an explicit patent grant, provided you carry the licence and notices with anything you pass on. That matters because it is a different file from the one governing the rest of the family — Qwen3.8-2.4T-A95B ships under a custom Qwen licence, not Apache 2.0 — so "it is a Qwen, so it is Apache" is not a safe assumption, and this is the member where it happens to be true.

Apache 2.0 on the weights is also the plain answer to the largest single qualified query this page sits next to, the uncensored and abliterated ones. What the licence permits is simple and worth saying without dressing: the published weights may be modified and redistributed, including modifications that change the model's behaviour, and derivative builds may be released under the same terms. That is the entire mechanism behind an abliterated release — a community or a lab publishes a modified copy of openly licensed weights and redistributes it. We run that route ourselves for this model: the Qwen3.8-27B-Uncensored-GGUF, NVFP4 and MLX builds are ours, published under the same Apache 2.0 terms as the original. What a hosted service will or will not generate is a separate question from what the licence allows, and this page does not answer it — the per-format uncensored pages cover the downloads.

What is not settled

Three things are genuinely open. The 1M-token context is a YaRN extension rather than a native window, and the model card itself cautions that static YaRN scaling can hurt short-input performance, so the practical ceiling for most local setups is the 262,144 native window and the extension is a per-deployment decision you have to test. Alibaba's own model card points readers seeking a managed service at Qwen Cloud, which it describes as coming soon for this model — so as of October 2026 the hosted tier the card advertises is not something we can confirm as generally available, and we are not going to describe its terms from a teaser. And the benchmark table on the card is Alibaba's own evaluation, unreproduced by the party that published it, which is a caveat that outlives the launch.

Availability

We serve this model. The identifier is qwen/qwen3.8-27b in our live model list — it is the dense multimodal 27B, it accepts text, images and video with a 262,144-position context window, and a reader can point an OpenAI-compatible client at it today. That is all we will say about the route here: no price claim, no throughput claim. If you are choosing on cost, work it out from your own traffic rather than from a number on this page.

The choice left to make is the format, and it is a hardware question before it is a software one. If you are running it on the machine in front of you, the GGUF route is where the quantisation tiers live and where you find out whether your card is enough. If you are on a Mac, the MLX route is the one that determines whether you have the unified memory at all. Read one of those two next, and come back to this page only if you want to remind yourself what the model is before you commit the disk space.