Qwen3.8 27B
Open-weight model from the Qwen team (Alibaba), released on 5 August 2026 under the Apache 2.0 license. It is a dense 27-billion-parameter model that also reads images and videos. Its native window is 262,144 tokens, stretchable to one million with a context-extension technique (YaRN). Thinking mode is on by default and is set per request: you can turn it off, or tune its depth with `reasoning_effort` (`low`, `medium`, `xhigh`). To run it at home, compressed GGUF builds range from 7.2 GB to 31.5 GB depending on quantization, and Ollama ships it as `qwen3.8:27b`. Qwen presents it as better than its 3.6 generation at code, professional work and long agentic tasks; those measurements are Qwen's own.
Strengths
- Apache 2.0 license: commercial use, modification and self-hosting with no in-house clause
- Fits on a single well-equipped machine: 16.6 GB in Q4_K_M, where large open models need a server
- Thinking adjustable on every request: you spend compute time only when the task deserves it
- Reads images and videos in the same model, without wiring a second tool
Limitations
- Dense model: all 27 billion parameters work on every token, so it runs slower on a laptop than models that only activate a fraction
- The published performance comparisons are Qwen's own, on its own tests: measure it on your case before handing it a pipeline
- Qwen itself warns that too little thinking can lengthen an agentic task: less analysis, more failures and retries
- Past 262,144 tokens you have to enable YaRN by hand, and the default one-million window is reserved for the announced hosted version
Best for
- Agent builders who want a local executor behind a proprietary planning model
- Teams that have to keep their data on their own machines and have at least 20 GB of free memory