Published: 17 August 2026
Alibaba released Qwen3.8-27B on 14 August 2026, an open weights dense model that accepts text, images and video and ships under the Apache 2.0 licence. Qwen3.8-27B has a 262,144 token native context window that Alibaba says extends to 1,000,000 tokens, and it is the smaller companion to Qwen3.8-Max, whose weights the Qwen team published two days earlier. On Alibaba’s own evaluations Qwen3.8-27B scores 61.7 percent on SWE-bench Pro against 53.4 percent for Claude Opus 4.6 Max.
What can Qwen3.8-27B do?
Qwen3.8-27B is a native vision language model rather than a text model with an adapter bolted on. The published weights accept text, images and video as input and return text, and the Qwen team ships runnable image and video examples against the local checkpoint in the model card. That matters because early coverage of the Qwen3.8 family described the open weights as text only, which was true of the Qwen3.8-Max checkpoint released on 12 August but is not true of Qwen3.8-27B. The configuration file for Qwen3.8-27B carries a full vision encoder and sets the language model only flag to false.
Alibaba positions the model at local agent work. In the release post the Qwen team says the model outperforms Qwen3.7-Plus overall and, in its words, shines in real world coding and office workflows. The practical claim is that a 27B dense checkpoint that fits on a single high end consumer or workstation GPU can now run the kind of long horizon tool calling that previously required a hosted frontier model. Qwen3.8-27B also ships in an FP8 variant, Qwen3.8-27B-FP8, which uses fine grained FP8 quantisation with a block size of 128 and which Alibaba says performs nearly identically to the original.
Qwen3.8-27B benchmarks and technical specs
Alibaba prints 27B as the parameter count for the language model. The published checkpoint totals 27,781,427,952 parameters in BF16 once the vision encoder is included, which is where the 27.78B figure circulating in coverage comes from. The model is dense, not a mixture of experts. It runs 64 layers at a hidden dimension of 5,120, in a repeating pattern of three Gated DeltaNet blocks to one Gated Attention block, and it was trained with multi token prediction. The context window is 262,144 tokens natively and Alibaba documents extension to 1,000,000 tokens through YaRN, with the caveat, stated in its own model card, that the common open source frameworks implement YaRN statically and can lose accuracy on shorter inputs as a result.
The benchmark table is Alibaba’s own, and the harness is disclosed: all models were run through the Claude Code harness at temperature 1.0, top_p 0.95 and a 256K context window, with the single exception of Claude Opus 4.6 Max on SWE-bench Pro, where Alibaba uses the officially reported score.
- On that table Qwen3.8-27B reaches 61.7 percent on SWE-bench Pro against 53.4 percent for Claude Opus 4.6 Max, 57.6 percent for Qwen3.7-Plus and 51.2 percent for Meta’s Muse Glimmer-30B.
- It scores 73.0 percent on Terminal Bench 2.1, behind Claude Opus 4.6 Max at 78.2 percent but well ahead of Muse Glimmer-30B at 51.7 percent.
- The largest generational jump is DeepSWE 1.1, where Qwen3.8-27B posts 42.2 percent against 13.3 percent for Qwen3.6-27B.
Two figures deserve a caveat. Alibaba’s widest margins sit on QwenSWEBench, where it reports 79.0 percent against 63.8 percent for Claude Opus 4.6 Max, and on CoWorkBench at 70.7 percent against 68.2 percent. Both are in house benchmarks with a one line description and no published task list or sample size, so they cannot be independently checked. Alibaba also states that it corrected problematic tasks in SWE-bench Pro and re-evaluated all baselines on the amended set, and that it fixed a small number of incorrect ground truth annotations in MathVision and CharXiv. Those are reasonable engineering decisions but they mean the scores are not directly comparable to numbers produced on the unmodified benchmarks.
On general reasoning the picture is more mixed than the coding results suggest.
- Qwen3.8-27B scores 89.2 percent on GPQA Diamond against 91.3 percent for Claude Opus 4.6 Max,
- 30.8 percent on Humanity’s Last Exam against 40.0 percent.
- It leads on LiveCodeBench v6 at 90.3 percent against 88.8 percent,
- IFBench at 79.5 percent against 62.5 percent.
- On the vision side the gaps are larger in Alibaba’s favour: 84.3 percent on OSWorld-Verified against 72.7 percent, 81.9 percent on AndroidWorld against 62.0 percent, and 90.0 percent on MathVision against 65.5 percent.
How does Qwen3.8-27B compare to Muse Glimmer and Qwen3.8-Max?
The direct comparison is Meta’s Muse Glimmer, a 30B open weights agentic model released on 10 August 2026 under the same Apache 2.0 licence and aimed at the same use case of an always on local agent. On the four benchmarks where Alibaba reports both, Qwen3.8-27B leads on all four: Terminal Bench 2.1 at 73.0 against 51.7, SWE-bench Pro at 61.7 against 51.2, IFBench at 79.5 against 77.0 and GPQA Diamond at 89.2 against 83.5. Those are Alibaba’s numbers for a competitor’s model, so treat the margins with the usual caution, but the harness and sampling parameters are published, which is more than most such comparisons offer.
Against its own family, Qwen3.8-27B is the deployable half of the Qwen3.8 release. Qwen3.8-Max is a 2.4 trillion parameter mixture of experts model with roughly 95 billion active parameters that Alibaba shipped on 3 August and whose open weights landed on 12 August. Those Max weights are text only. Qwen3.8-27B is roughly a hundredth of the size, keeps the vision and video input, and is the checkpoint most teams will actually be able to run. Alibaba pre-announced it in the Qwen3.8-Max launch post on 3 August, saying the 27B was also going open weights, and delivered eleven days later.
Qwen3.8-27B availability, licence and pricing
The weights are on Hugging Face as Qwen/Qwen3.8-27B and Qwen/Qwen3.8-27B-FP8, and on ModelScope. Both carry Apache 2.0, confirmed in the licence file rather than only in the model card front matter, which means commercial use is permitted without a revenue share clause. That is a meaningful difference from Qwen3.8-Max, which shipped under a bespoke licence. Serving recipes are published for SGLang and vLLM.
First party hosted pricing does not exist yet. The model card says the Qwen Cloud service is coming soon, and the Qwen Cloud page for the model returned a 404 as of 17 August 2026. For teams that want a hosted endpoint today the route is OpenRouter, which lists Qwen3.8-27B at 0.45 dollars per million input tokens and 3.20 dollars per million output tokens with a 262,144 token context window. Anyone quoting an official Alibaba API price for this model is quoting something Alibaba has not published.
For the full details on Qwen3.8 please visit the official Qwen3.8 model announcement page.
