Qwen3.8-27B Open-Sourced: 27B Parameters, Outperforms Claude Opus 4.6 Max on Multiple Benchmarks

Qwen3.8-27B was officially open-sourced on August 15, 2026, with a total of 27 billion parameters, a version long awaited by the community. Its key selling point is that after quantization, it can be fully deployed on a 24GB VRAM RTX 3090 or 4090, significantly lowering the barrier for local deployment.

The model natively supports a 262K token context, expandable to 1 million tokens, while retaining native multimodal capabilities for processing images, PDFs, charts, and even videos. These features ensure that despite its smaller size, it does not sacrifice long-context or multimodal abilities.

In official benchmarks, Qwen3.8-27B surpasses Claude Opus 4.6 Max on multiple coding and agent tasks. Specifically, it leads by 8.3 points on SWE-bench Pro and by 15.2 points on QwenSWEBench; on CoWorkBench it scores 70.7, exceeding Opus 4.6 Max's 68.2.

In multimodal and operation evaluations, it scores 84.3 on OSWorld-Verified (Opus 4.6 Max: 72.7), 81.9 on AndroidWorld (Opus 4.6 Max: 62.0), and improves WebArena-Verified from 48.8 in the previous generation to 64.8. With CI enabled, it achieves 94.6 on visual math problem solving and 85.6 on general visual reasoning (65.7 without CI).

These figures come from official benchmarks and are source claims, not yet independently verified by third parties. The limitations of benchmarks mean the 'Opus-level' claim should be treated with caution.

Qwen3.8-27B uses a 64-layer architecture, with 48 layers employing Gated DeltaNet linear attention and 16 layers retaining full attention, arranged in a cycle of 'three linear attention layers plus one full attention layer.' This design reduces computational and cache pressure for long sequences, supporting the 262K native context.

The model includes a built-in reasoning effort setting with three levels—xhigh, medium, and low—allowing users to adjust reasoning depth to balance speed and cost; the Thinking mode can be disabled. Additionally, preserve_thinking is enabled by default, allowing the agent to retain earlier reasoning steps during long tasks, improving sequential decision-making efficiency.

Community users have showcased various applications, such as pixel-style pagodas, Tetris, and Snake, with one user remarking, 'It feels like having an Opus-level agent.' Another user deployed Qwen3.8-27B-FP8 on a single NVIDIA GH200, handling 10 real requests simultaneously, each with a maximum output of 16K tokens and a 262K context. The first streaming token returned in under 10ms, and all 10 requests completed successfully.

In multimodal tests, a user had the model identify 96 timestamped events from an 11-minute 1935 film, completing the task in 157 seconds with a time error of about 2 seconds, all on a single GPU. These are community reports and are source claims, not officially confirmed.

Official support includes Transformers, vLLM, SGLang, and TokenSpeed. For production environments, SGLang and vLLM are recommended; local users can deploy via Hugging Face quantized versions.

Qwen3.8-27B achieves a balance among parameter size, context length, and multimodal capabilities, with impressive benchmark results. However, the 'Opus-level' claim is primarily based on official data and lacks independent verification. Actual performance may vary depending on task type and deployment environment.

Local deployment is feasible, but quantization may introduce precision loss, and high-concurrency scenarios still require high-end hardware such as the GH200. The practical effects of features like reasoning effort and preserve_thinking await more user feedback.

Overall, Qwen3.8-27B offers developers the possibility of running a powerful agent on consumer-grade hardware, but its 'Opus-level' claim should be viewed as a relative performance under vendor benchmarks, not an absolute equivalence.