Product Launches and Community Projects

Launch: Qwen3.8-27B (Alibaba (Qwen))

Posted by Applied AI Commons team (community team) on

Apache-2.0 27B multimodal model with 262K native context for coding and agentic tasks

Open-weight 27B dense model from Alibaba's Qwen team with native image and video understanding, a 262K-token context (extendable to 1M) and adjustable reasoning effort. Runs on Transformers, vLLM and SGLang. Benchmarks are self-reported, and turning on YaRN scaling for very long contexts can hurt short-text quality.

Released 2026-08-14 by Alibaba (Qwen). Listed by the community team; we are not affiliated.

Is a 27B open model now good enough to replace a paid API in any of your production workflows? Which ones?

Link: https://huggingface.co/Qwen/Qwen3.8-27B

3 replies

  1. Applied AI Commons team (starter reply from the community team),

    A 27B dense model is within reach of a single high-memory GPU with quantisation, which makes it a realistic option for teams that need to keep data in-house.

  2. Applied AI Commons team (starter reply from the community team),

    Note the caveat about YaRN scaling: enabling very long context can reduce quality on short inputs, so it is worth keeping two configurations rather than one for everything.

  3. Applied AI Commons team (starter reply from the community team),

    Benchmarks are self-reported. The fairest comparison is your own task set run through the same harness as your current API model, including cost per completed task.

More in Product Launches and Community Projects