Unsloth has released GGUF files for Qwen3.8-27B. The company stated that its "Unsloth Dynamic 3.0" achieves up to a 10% or more improvement in top-1% accuracy compared to other major quantization providers at equivalent sizes.
Qwen3.8 is positioned as the highest-performance generation within the Qwen open-weight model family. Based on the Qwen3.5 architecture, it achieves significant performance gains in coding, professional tasks, research, and long-term agent tasks.
Qwen3.8-27B is a native vision-language model capable of understanding images and videos. It is designed with flexible thought-control capabilities to complete complex multi-step tasks more reliably and is provided as a compact, easy-to-deploy dense model.
To obtain optimal performance, specific sampling parameters and output length settings are recommended. Particularly for agent tasks, ensuring sufficient output length is critical for generating detailed responses. By appropriately setting token limits for internal reasoning and final output within the 1M context length, it is possible to balance complex reasoning with high-quality deliverables.
For processing ultra-long texts, it natively supports a context length of up to 262,144 tokens. For long-term tasks exceeding this limit, the use of RoPE scaling techniques such as YaRN is recommended. To improve video understanding efficiency, setting the longest_edge parameter in video_preprocessor_config.json to 469,762,048 enables high-frame-rate sampling even for hour-long videos.
Source: Unsloth Qwen3.8-27B GGUF files (HN 69pt, 5 comments) (HN Search (backfill), 2026-08-15)