Qwen3.8-27B: Alibaba's New Open-Weight Model That Punches Above Its Weight
Alibaba has released the open weights of Qwen3.8-27B, a 27-billion-parameter dense multimodal model under the Apache 2.0 license. Built to run on local hardware, it dares to compete with models several times its size on coding and office work, and it wins on many benchmarks.
What is Qwen3.8-27B
Qwen3.8-27B is the compact member of the Qwen3.8 generation, announced on August 14, 2026. It is a dense model (no mixture of experts) with native text, image and video understanding, built on the architectural foundation of Qwen3.5 with a hybrid mix of Gated DeltaNet linear attention and classic attention, plus multi-token prediction. The combination targets maximum efficiency on a single GPU without giving up quality.
The release came together with open weights for Qwen3.8-2.4T-A95B, the Max-class mixture-of-experts model aimed at heavier agentic workloads. With one move, Alibaba covers both ends of the market: a lightweight model for local deployment and a frontier-scale one with open weights from day one.
Key features
Native multimodality
The model understands images and videos up to an hour long: scientific diagrams, documents, screenshots and real-world scenes. On visual benchmarks it scores 90.0 on MathVision (94.6 with a fixed answer format) and 91.1 on OmniDocBench 1.5, ahead of its predecessor Qwen3.6-27B.
262K context, expandable to 1M
The native window is 262,144 tokens and can be extended to one million using YaRN scaling. Alibaba recommends tuning the scaling factor to the typical length of each use case, since static YaRN can slightly hurt performance on short texts.
Flexible thinking control
Thinking mode is enabled by default and can be turned off per request. The reasoning_effort parameter adjusts depth (xhigh, medium, low), and preserve_thinking keeps the reasoning trace across messages, which is especially useful for agents and also improves KV cache usage.
Benchmark results
Versus Muse Glimmer-30B
The most discussed comparison is with Muse Glimmer, Meta's 30B local model released days earlier. On every test where both models ran, Qwen3.8-27B wins: Terminal Bench 2.1 (73.0 vs 51.7), IFBench (79.5 vs 77.0) and GPQA Diamond (89.2 vs 83.5). Meta did not report results on harder benchmarks such as DeepSWE 1.1, JobBench, NL2Repo-Bench or LiveCodeBench v6, which limits the overall comparison.
Versus much larger models
Against Qwen3.7-Plus, a bigger proprietary model, the 27B wins most rows: SWE-bench Pro 61.7 vs 57.6, QwenSWEBench 79.0 vs 59.2 and Agents' Last Exam (20.4 vs 13.2 Pass@1). Against Opus4.6 Max the picture is mixed: it wins on SWE-bench Pro, IFBench (79.5 vs 62.5) and LiveCodeBench v6 (90.3 vs 88.8), but trails on Terminal Bench 2.1 and scientific reasoning.
The strong suit: agentic office work
The clearest jump over its predecessor is in long-horizon office tasks. On CoWorkBench it scores 70.7, ahead of Opus4.6 Max (68.2) and far above Qwen3.6-27B (61.0). On JobBench, which measures complete professional tasks, it climbs to 33.4 versus 21.8 for the previous generation, roughly 50 percent higher.
Availability and ecosystem
Weights are on Hugging Face and ModelScope under Apache 2.0, with GGUF builds prepared by Unsloth for local use. The model works with Transformers, vLLM, SGLang and TokenSpeed, and Qwen Cloud is preparing a hosted version with a 1M context by default and built-in tools.
Conclusion
Qwen3.8-27B proves that size is not everything: with 27B parameters and an efficient hybrid architecture, it outperforms models several times larger on coding and office automation, and its open weights make it a very attractive option for running locally. If you want to try cutting-edge models on your own hardware, this is a strong starting point.
