Alibaba's Qwen3.8 27B scores 52 on Artificial Analysis, topping open-weight peers

Alibaba has released Qwen3.8 27B, a 27-billion-parameter open-weights model that lands a 52 on Artificial Analysis’s composite Intelligence Index — a benchmark spanning reasoning, knowledge, math, and coding. That figure sits far above the quoted median of 9 for comparable open-weight models in its size class, positioning it as a standout among self-hostable options released this month (its launch date is August 14, 2026). The model is multimodal on the input side, accepting both text and images while producing text output, and it is a reasoning model that uses chain-of-thought before answering.

The practical draw is the licensing and cost profile. Qwen3.8 27B ships under Apache 2.0, so the weights are downloadable, self-hostable, and cleared for commercial use — which is why Artificial Analysis lists its input and output pricing as $0.00 rather than a real hosted API rate. Buyers evaluating it should read that zero as “bring your own infrastructure,” not free inference. Its long context window is a further selling point, though the page reports it inconsistently, citing 256k tokens in the summary and 260k in the FAQ.

One caveat worth flagging: the model is notably verbose. Running the Intelligence Index, it generated 160M output tokens against a class median of 43M, meaning the strong score comes with roughly 3.7x the token spend of typical peers — a real cost consideration for anyone self-hosting or paying per token downstream. For a reasoning model that trades extra thinking for accuracy, that verbosity is expected, but it tempers the headline efficiency story.