Alibaba drops Qwen 3.8 27B with local FP8 support

Alibaba's Qwen team released Qwen 3.8 27B, bringing high-end coding and vision capabilities to local hardware.

The release includes official FP8-quantized weights on Hugging Face that run directly in vLLM, SGLang, and TokenSpeed. The model features native image and video understanding, a 262k context window extensible to 1 million tokens, and built-in thinking controls.

Why it matters: Mid-sized open models usually fall short on long agentic tasks. On benchmarks like SWE-bench Pro, Qwen 3.8 27B scores 61.7—beating Muse Glimmer-30B's 51.2—while remaining compact enough to run on local GPUs.

Here's the gist on controls: thinking mode is on by default, but you can disable it per request or adjust depth using the reasoning_effort parameter.

Cloud GPUs remain expensive, but local weights just got a lot smarter.