Alibaba’s Qwen team has published Qwen3.8-27B as an open-weight download, and the release is unusually easy to verify from primary sources. The Hugging Face API metadata for the repository records a lastModified timestamp of 2026-08-14T15:00:01Z, a license field reading apache-2.0, and "gated":false, meaning the weights can be pulled without applying for access first. The same metadata puts the on-disk weight count at 27,781,427,952 parameters in BF16, a figure that includes the model’s vision encoder.
The same day, AMD announced local-inference support for the model on its own processors and graphics cards. “Today, AMD is delivering Day 0 support for Qwen3.8 27B, giving developers a path to run this state-of-the-art dense model locally on AMD-powered PCs and workstations from the moment it becomes available,” AMD’s blog post reads. The post is credited to Syed Muhammad Usman Pirzada, a product marketing manager for client AI, and Slobodan Josic. Qwen’s model card carries no individual byline and is written in the team’s collective voice.
This is a follow-up rather than a fresh thread. On 4 August this site reported on the Qwen3.8 family and noted that the 27B was still forthcoming. The 27B is now the part that has actually shipped.
What Is Actually in the Repository
Qwen’s own model card lists the language-model side at “Number of Parameters: 27B,” the smaller of the two counts and the one to use against other language models. The API’s 27.78B total is the same model measured with the vision encoder included. Both are correct and they are not interchangeable.
Context length is the specification most likely to change how the model gets used: “Context Length: 262,144 natively and extensible up to 1,000,000 tokens.” That is the language model’s window, and video shares it. The card recommends setting longest_edge to 469,762,048 in the video preprocessor config, which it says corresponds to 224k video tokens, so a long video can consume most of a 262,144-token context on its own.
The architecture is a hybrid. The card gives “Number of Layers: 64” and a “Hidden Layout” of “16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN)).” Gated DeltaNet is the linear-attention component, Gated Attention the full-attention one. Multiplying out the stated repeating unit gives 48 of the former and 16 of the latter, which is arithmetic on Qwen’s pattern, not a split Qwen prints on the card.
Vision is native rather than bolted on. Per the card, “Qwen3.8-27B brings these advances to 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.” The card frames the generation in its own collective voice: “Following the widespread community adoption of the Qwen3.5 and Qwen3.6 series, we are pleased to introduce Qwen3.8, the most capable generation in the Qwen open-model family to date.”
One date detail: the API also carries a createdAt value of 2026-08-05T08:22:59Z, nine days earlier. The repository was not created on 14 August, it was last updated then, and that timestamp is the one matching the weights and README now live.
AMD’s Numbers, and What They Are Not
AMD published throughput figures alongside its support announcement: “Early testing shows strong local performance for Qwen3.8 27B on AMD hardware, reaching up to 24.5 tokens per second on AMD Ryzen™ AI Max+ 395 and up to 51.8 tokens per second on a single AMD Radeon™ AI PRO R9700.”
AMD’s framing matters as much as the numbers. The same post describes them as preliminary and specifies the conditions: “These preliminary results were measured on Windows using the popular llama.cpp project with the Vulkan backend, with MTP=4 on Ryzen AI Max+ 395 and MTP=2 on Radeon AI PRO R9700, using average token-generation throughput across three or more runs.” That is a vendor measuring its own silicon, on one operating system, through one inference stack, with different multi-token-prediction settings on each part. Read them as a ballpark for AMD hardware, not a neutral benchmark.
On memory, AMD states: “Note that this model requires roughly 24GB of variable graphics memory (VGM) or VRAM to run comfortably and will also run on older, supported AMD platforms in LM Studio.” That is AMD’s number for comfortable operation via LM Studio and llama.cpp on AMD hardware, and it has not been checked against other vendors’ cards or other serving stacks.
This is where the earlier coverage needs an honest note. The 4 August piece on this site carried a lower, pre-release memory figure for the then-unreleased 27B, attributed to a post by Daniel Han of Unsloth. That piece named no quantization and no inference stack, so the two numbers cannot be compared directly and neither has been shown to be wrong. What can be said now is narrower and more useful: AMD’s roughly 24GB is the figure a hardware vendor has published for its own stack, on shipping weights, and it is the one to plan around.
The Benchmark Table Cuts Both Ways
Qwen’s card includes two comparison tables, and reading only the winning rows misrepresents them. On SWE-bench Pro, Qwen3.8-27B scores 61.7, the best of the five models compared, ahead of Qwen3.6-27B at 53.5, Qwen3.7-Plus at 57.6, Muse Glimmer-30B at 51.2 and Opus4.6 Max at 53.4. It also leads on QwenSWEBench (79.0 against 63.8), LiveCodeBench v6 (90.3 against 88.8), CoWorkBench (70.7 against 68.2) and IFBench (79.5 against 62.5).
That SWE-bench Pro row carries a footnote from Qwen: “Except for Opus4.6 Max, which uses the officially reported score, all models are evaluated with the Claude Code harness at temp=1.0, top_p=0.95, and a 256K context window.” The Opus figure was not run on Qwen’s harness, which makes that row a mixed-methodology comparison by Qwen’s own admission. The footnote is scoped to SWE-bench Pro, not to the table as a whole.
Counting the rows where Opus4.6 Max carries a score, Qwen3.8-27B trails it on exactly four, and two of them are coding rows rather than reasoning rows. On agentic terminal coding, Terminal-Bench 2.1 (Terminus), it scores 73.0 against 78.2 for Opus. On repo-level code generation, NL2Repo-Bench, it scores 42.3 against 47.6. On GPQA Diamond, graduate-level science questions, it scores 89.2 against 91.3. On HLE, multidisciplinary reasoning judged by GPT-4o per Qwen’s own footnote, the gap is widest: 30.8 against 40.0.
The multimodal table runs the other way throughout. On computer use, OSWorld-Verified, Qwen3.8-27B scores 84.3 against 72.7 for Opus4.6 Max; on mobile use, AndroidWorld, 81.9 against 62.0; on multimodal software engineering, SWE-MM, 38.6 against 27.1. Opus4.6 Max carries no score at all on several rows in that table.
What This Means
A 27B dense model with a 262,144-token native window, native image and video understanding, and no download gate is now available under Apache 2.0, and a chipmaker published its own local-inference numbers on the same day. For anyone with a card in the roughly 24GB class AMD describes, that is testable this week rather than a claim to wait out.
The benchmark story is narrower than a headline would make it. On Qwen’s own tables the model leads Opus4.6 Max across the multimodal rows and most agentic ones, and trails it on four: Terminal-Bench 2.1, NL2Repo-Bench, GPQA Diamond and HLE. Two of those four are coding rows, so “wins at code, loses at reasoning” is too tidy a summary of a table that is genuinely mixed.
The memory question is answerable in one direction only. AMD’s roughly 24GB is a vendor figure for LM Studio and llama.cpp on AMD hardware, while the lower figure this site carried before release came with no stated quantization or stack behind it. Both can stand, but only one is attached to a vendor and to shipping weights.
The Bottom Line
Ungated Apache 2.0 weights, a 262,144-token native context extensible to 1,000,000, native vision-language, and AMD publishing preliminary throughput of up to 24.5 and 51.8 tokens per second on its own parts the same day. The number to plan hardware around is AMD’s roughly 24GB for its own stack, not the lower pre-release figure this site carried before the weights existed.