Qwen Releases FP8-Quantized 27B Model on Hugging Face

The Qwen Team has officially released Qwen3.8-27B-FP8, a fine-grained FP8 quantized variant of its latest open-source native vision-language model. Published in August 2026 on Hugging Face, this deployment-friendly dense architecture delivers frontier-level capabilities in coding, agentic workflows, and long-horizon task execution while maintaining nearly identical performance to its original precision counterpart. The release follows widespread community adoption of the Qwen3.5 and Qwen3.6 series and builds directly upon the Qwen3.5 architectural foundation. The FP8 variant utilizes a block size of 128 to achieve substantial memory and compute efficiency without sacrificing accuracy. Model weights and configuration files are optimized for seamless integration with major inference frameworks, including Hugging Face Transformers, vLLM, SGLang, and TokenSpeed. As a native vision-language architecture, the 27B model natively processes images and videos alongside text, featuring flexible thinking controls designed to improve reliability across complex, multi-step operations. Independent benchmark evaluations demonstrate significant advancements across technical and creative domains. In software engineering and agentic coding, the model achieves a 78.2 score on Terminal Bench 2.1 and 61.7 on SWE-bench Pro, outperforming comparable architectures in repo-level generation and autonomous debugging. Long-horizon productivity tasks show similar gains, with a 70.7 score on CoWorkBench and 33.4 on JobBench. Multimodal capabilities remain equally competitive, recording an 84.3 on OSWorld-Verified for computer use, 81.9 on AndroidWorld for mobile automation, and 90.2 on CharXiv for scientific chart analysis under controlled inference conditions. Deployment workflows support both local and cloud infrastructure. The model operates in a default thinking mode, generating explicit reasoning traces before delivering final outputs, though users may disable this feature via standard API parameters. Sampling recommendations specify a temperature of 1.0 with a top_p of 0.95 for reasoning tasks, and 0.7 with 0.80 for direct instruction modes. The architecture natively supports a 262,144-token context window, with recommended RoPE scaling techniques like YaRN enabling extended processing for ultra-long documents and hour-scale video analysis. For enterprises prioritizing managed inference, Qwen Cloud will soon host the Qwen3.8-27B variant with production-grade enhancements, including a default 1-million-token context length and integrated tooling. The release underscores a strategic shift toward compact, high-efficiency dense models that rival larger parameter architectures in specialized agentic and multimodal applications, offering developers a streamlined pathway to deploy frontier reasoning capabilities without prohibitive infrastructure costs.