A 27B laptop model scores like a rented one, and thinks three times as hard to do it

Alibaba published the weights for Qwen3.8-27B on Hugging Face on Friday under an Apache 2.0 licence ^[1], and on Monday the benchmark firm Artificial Analysis scored it 52 on its Intelligence Index, the same number it assigns OpenAI's GPT-5.6 Luna at its highest reasoning setting ^[2][3]. For teams paying per token for cloud inference, the consequential detail is not the score but the footprint: a compressed 4-bit build of the model is roughly 17GB ^[4].

The licence does more work here than the parameter count. Apache 2.0 lets companies inspect, change and host the model themselves ^[5], which is what turns a benchmark line into a procurement question. Artificial Analysis, cited by the South China Morning Post, put the 27B first in a class of 135 on a composite of nine coding, science and reasoning tests ^[6][7], close to DeepSeek's 1.7-trillion-parameter V4-Pro and Zhipu's 753-billion-parameter GLM-5.2 ^[8][9] - models roughly 63x and 28x its size ^[10][11]. Alibaba's own table claimed 61.7 on SWE-bench Pro and 90.3 on LiveCodeBench, topping a listed Claude Opus 4.6 result ^[12][13]. VentureBeat noted that some evaluations were internal and the test setups were not identical, making the numbers poor grounds for declaring a winner ^[14]. The independent number is the one to plan against.

The hardware claim survives contact with a real desk. Full precision needs about 56GB of GPU memory ^[15]; Simon Willison ran a roughly 17GB version on an Apple laptop and an Nvidia desktop, where it wrote code, read images and ran a coding-agent loop ^[16]. Cline, the open-source coding tool, said on X that this was "the first time a local model has scored frontier model capability" and reported 51 on its own agentic test, above Claude Opus 4.8 at maximum reasoning ^[17][18].

Then the bill arrives in a different currency. Artificial Analysis measured 160 million output tokens across its testing against a 43-million median for comparable open-weight models ^[19], roughly 3.7 times as many ^[20]. Willison saw the same thing at human scale: a request to draw a simple image took 21 minutes and more than 22,000 reasoning tokens because the model defaults to maximum reasoning effort, and he recommends turning that down for ordinary local use ^[21]. That asymmetry is the actual build-vs-rent argument. Rented, a model that emits 3.7x the tokens partly gives back what its size saves; owned, the marginal token is electricity and wall-clock time, and reasoning effort becomes a dial you control rather than a line item. Willison reported about a 72 percent performance gain on his Nvidia machine after enabling Multi-Token Prediction ^[22]. The investor Tomasz Tunguz found a similar trade-off testing the model in his own agent stack ^[23].

Demand figures are unsettled: The Information reported more than a million downloads in a few days and called it one of Alibaba's fastest-growing models, while Cybernews reported 3 million Hugging Face downloads in the first three days ^[24][25].

Watch whether inference software closes the latency gap, since that is what makes a local agent loop usable rather than impressive. Watch the default reasoning setting, which currently costs minutes on trivial tasks. And watch Alibaba, which is courting on-device users who skip the paid cloud service while charging its largest customers to run other models ^[26][27].