Even as Chinese models are closing in on the AI frontier, they’re also making moves in the local models space.
Alibaba has released Qwen3.8-27B, a dense multimodal model that the company is positioning as its answer to the growing crowd of local-first, single-GPU-friendly models from rivals like Meta and Google. The release comes packaged with a promise Alibaba made weeks ago and has now delivered on: open weights, right out of the gate, under the Apache 2.0 license. Alongside the 27B model, Alibaba has also opened weights for Qwen3.8-2.4T-A95B, the Max-class version of the same generation, giving developers a choice between a lightweight model that runs comfortably on local hardware and a frontier-scale one built for heavier agentic workloads.
Qwen3.8-27B ships with a native context window of 262K tokens, extendable up to 1 million via YaRN, and Alibaba is billing it as a model built specifically “for builders” — efficient enough to run locally while still competing with models several times its size on coding and office-work benchmarks. The company’s own numbers show it not just holding its own against Qwen3.7-Plus, a larger proprietary model from the same family, but outperforming it on several fronts.
The comparison that will likely draw the most attention, though, is with Meta’s Muse Glimmer, a 30B local model that Meta released just days ago and had already positioned as beating Qwen3.6-27B on most of its tests. Qwen3.8-27B appears to have been built, at least in part, as a direct response.
The Benchmarks
Alibaba tested Qwen3.8-27B against four models: its own predecessor Qwen3.6-27B, the larger Qwen3.7-Plus, Meta’s Muse Glimmer-30B, and Anthropic’s Opus4.6 Max, across coding, agentic, and general reasoning categories.
The gap over Muse Glimmer is where the release makes its strongest case. On Terminal Bench 2.1, which measures agentic terminal coding, Qwen3.8-27B scores 73.0 against Muse Glimmer’s 51.7, a lead of more than 20 points. On IFBench, a test of instruction following, Qwen posts 79.5 to Muse Glimmer’s 77.0. On GPQA Diamond, a scientific reasoning benchmark, Qwen edges ahead at 89.2 versus 83.5. Muse Glimmer’s results are missing entirely from several of the harder benchmarks Alibaba ran, including NL2Repo-Bench, DeepSWE 1.1, JobBench, and LiveCodeBench v6, which limits how directly the two models can be compared across the board, but on every test where both models were run, Qwen3.8-27B comes out on top.
The more interesting story, though, is how Qwen3.8-27B stacks up against models many times its size. Against Qwen3.7-Plus, a model Alibaba has not disclosed the parameter count of but which sits a tier above in its own lineup, the 27B model wins on most rows. It posts 61.7 on SWE-bench Pro against Qwen3.7-Plus’s 57.6, and takes QwenSWEBench by a wide margin at 79.0 versus 59.2. It also leads on Agents’ Last Exam, a benchmark built to test frontier agentic reasoning, with a Pass@1 score of 20.4 and a raw score of 42.9, both comfortably ahead of Qwen3.7-Plus’s 13.2 and 33.6.
Against Opus4.6 Max, the picture is more mixed, which is expected given the difference in scale and intended use case. Opus4.6 Max stays ahead on Terminal Bench 2.1 (78.2 vs 73.0), repo-level code generation on NL2Repo-Bench (47.6 vs 42.3), and scientific and multidisciplinary reasoning on GPQA Diamond and HLE. But Qwen3.8-27B actually beats it on SWE-bench Pro (61.7 vs 53.4), QwenSWEBench (79.0 vs 63.8), instruction following on IFBench (79.5 vs 62.5), and competitive coding on LiveCodeBench v6 (90.3 vs 88.8). For a model built to run locally, beating a frontier-scale proprietary model on real-world coding and instruction-following tasks is a notable result, even if it trails on the harder reasoning benchmarks.
The category where Qwen3.8-27B separates itself most clearly is agentic office work, an area Alibaba has been investing in heavily across its recent releases. On CoWorkBench, its in-house benchmark for long-horizon productivity tasks, the model scores 70.7, ahead of Opus4.6 Max’s 68.2 and well clear of Qwen3.6-27B’s 61.0. On JobBench, which tests professional job-task completion, it posts 33.4, roughly 50 percent higher than Qwen3.6-27B’s 21.8.
Open Weights, As Promised
The headline framing from Alibaba’s Qwen team was less about any single benchmark and more about following through on a commitment. The company had said open weights for Qwen3.8 were coming, and this release delivers on both ends of that promise at once — a compact 27B dense model for local and edge deployment, and the much larger 2.4T-parameter A95B mixture-of-experts model for teams building serious agentic infrastructure.
That dual release mirrors a pattern Alibaba has leaned on before. Smaller and mid-sized Qwen models have generally shipped open from day one, while the largest Max-class models have typically launched closed and proprietary before an eventual open release, if one comes at all. Qwen3.8-27B fits the former category cleanly, and its arrival alongside the open Max-tier weights suggests Alibaba is trying to cover both ends of the market in one move, rather than staggering the releases the way it has with past generations.