Qwen3.8-27B Weights Officially Released: Dense 27B Model Scores 61.7 on SWE-bench Pro, Delivering the Main Event for Single-Machine Users

Qwen3.8-27B has finally gone from “available next week” to downloadable. Its official Hugging Face page lists an Apache 2.0 license, approximately 28B parameters, and BF16 weights. Architecturally, it is a causal language model with a vision encoder. It natively supports a 262,144-token context window, extendable to 1,000,000 tokens through long-context techniques. These specifications make its positioning clear: Qwen3.8-Max is the cloud flagship, while the 27B model is for users who want control over the weights, costs, and data boundaries.

The standout official results are in software engineering. Qwen3.8-27B scores 73.0 on Terminal Bench 2.1, 61.7 on SWE-bench Pro, 42.3 on NL2Repo-Bench, 42.2 on DeepSWE 1.1, and 79.0 on QwenSWEBench. By comparison, Qwen3.6-27B scores 63.4, 53.5, 36.2, 13.3, and 49.3, respectively. This does not look like a minor update. The gains on DeepSWE and QwenSWEBench in particular suggest that the Qwen team is specifically targeting long-horizon coding agents: reading repositories, operating terminals, fixing errors, and responding to environmental feedback—not merely optimizing scores on isolated coding problems.

Its general reasoning and agent-task results are also notable. It scores 90.3 on LiveCodeBench v6, up from 83.9 for Qwen3.6-27B and slightly above the 88.8 listed for Opus4.6 Max in the table. It also scores 89.2 on GPQA Diamond, 30.8 on HLE, and 79.5 on IFBench. For long-running workflows, it reaches 70.7 on CoWorkBench, 33.4 on JobBench, and 20.4 Pass@1／42.9 Score on Agents' Last Exam—all clear improvements over the previous-generation 27B model. Together, these figures point to one conclusion: the 27B class is no longer positioned merely as a local chat model, but increasingly as a workforce-replacement layer capable of tool calling and multistep tasks.

The multimodal scores are equally noteworthy: 84.3 on OSWorld-Verified, 64.8 on WebArena-Verified, 81.9 on AndroidWorld, 47.1 on RecreationBench, 38.6 on SWE-MM, and 62.9 on Vision2Web. For MathVision, the official results list 90.0 without CI and 94.6 with CI; on BabyVision, the score rises from 65.7 without CI to 85.6 with CI. Here, “with CI” does not mean confidence interval—it refers to an auxiliary condition in the benchmark configuration. The two should not be confused when interpreting the results.

To put it fairly, this model card makes Qwen3.8-27B highly appealing, but not yet an automatic replacement for every local model. SWE-bench Pro, NL2Repo, DeepSWE, and Vision2Web largely use a Claude Code harness or internal Qwen benchmarks. Some benchmarks also note corrected tasks or the use of GPT-4o / GPT-5.4 as the judge. This does not make the scores untrustworthy; it is a reminder that official results are useful for deciding whether a model deserves evaluation, not as a direct guarantee of production performance. The most sensible approach is to test it across your own repository repair, terminal-task, UI-reconstruction, document-understanding, and long-context workflows, comparing it over multiple runs on identical tasks against your current Qwen3.6-27B, GLM, Kimi, or proprietary APIs.

If Qwen3.8-Max tells the market where Qwen’s performance ceiling lies, Qwen3.8-27B is more like a work light switching on in the dark. It is not the largest model, but it may be the one that the greatest number of engineering teams actually experiment with, quantize, deploy, and fine-tune. For local AI users, today’s news is about more than a model release: it means the 27B sweet spot is worth getting excited about again.