Alibaba’s new small model runs on a laptop and scores like a cloud one

Alibaba has released a small AI model that runs on a personal computer and scores like a cloud one. The model, Qwen3.8-27B, has been downloaded more than a million times in a few days, The Information reported. Juro Osawa wrote that it is one of the company’s fastest-growing models to date.

The model is open source and free to download. It has 27 billion parameters, and it understands text, image and video prompts, according to The Information. Alibaba published the weights on Hugging Face on Friday under an Apache 2.0 licence, VentureBeat reported. That licence lets companies inspect, change and host the model themselves.

The interest tracks a wider shift. Alibaba is chasing the market for small models that run on a user’s own hardware. Those models skip the paid cloud service. The 27B is part of the same family as Qwen3.8-Max, the large flagship. It also sits alongside the models Alibaba recently began charging its biggest users to run.

It matches models many times its size

The attention comes from the benchmarks. Qwen3.8-27B matched much larger near-frontier rivals while running on everyday hardware, the South China Morning Post reported. Xinmei Shen wrote that the model performed on par with OpenAI’s GPT-5.6 Luna. The US lab had billed Luna as the most cost-efficient model in its latest flagship series.

It also came close to two large Chinese open-weight models. It nearly matched DeepSeek’s V4-Pro, released last week with 1.7 trillion parameters. It also came close to Zhipu’s GLM-5.2, a 753 billion-parameter model from June, the SCMP said. Those figures came from the benchmark firm Artificial Analysis.

The independent scores landed on Monday. Artificial Analysis gave Qwen3.8-27B a score of 52 on its Intelligence Index. That index is a composite of nine tests, spanning coding, science and reasoning. That is the same score the firm assigns GPT-5.6 Luna at its highest reasoning setting. The index ranks the model first in its class of 135.

Alibaba’s own launch numbers set the tone first. The company reported scores of 61.7 on SWE-bench Pro and 90.3 on a coding test called LiveCodeBench, VentureBeat said. In Alibaba’s published table, the 27B model even topped a listed Claude Opus 4.6 result on those two. VentureBeat noted that some of the evaluations were internal, and that the test setups were not identical, so the numbers were poor grounds for declaring a winner.

Developers took notice

The comparison is what drew developers. “This is the first time a local model has scored frontier model capability,” the open-source coding tool Cline wrote on X. On the firm’s separate agentic test, the model scored 51. That beat Claude Opus 4.8 at maximum reasoning, a frontier model Anthropic released less than three months ago.

The point is the hardware, not the leaderboard

The model’s size is the story. Running it at full precision needs about 56GB of GPU memory, VentureBeat reported. A compressed 4-bit version shrinks the file to roughly 17GB. That puts it within reach of a high-end gaming desktop or a well-equipped laptop.

One developer put that to the test. Simon Willison ran a roughly 17GB version on an Apple laptop and an Nvidia desktop, he wrote. It wrote code, read images and ran a coding-agent loop. “The fact that a 17GB file can do all of this stuff on my home machines is a miracle,” Willison wrote.

Usage climbed fast. Qwen3.8-27B passed 3 million Hugging Face downloads in its first three days, Cybernews reported. That is a larger figure than the million-plus The Information cited over a similar window. Either way, the downloads piled up fast.

The catch is that it thinks a lot

The performance comes at a cost in speed. Qwen3.8-27B buys some of its quality by generating far more reasoning text than its rivals, VentureBeat reported. Artificial Analysis said the model produced 160 million output tokens across its testing. The median for comparable open-weight models is 43 million.

Willison hit the same wall. A request to draw a simple image took 21 minutes and more than 22,000 reasoning tokens, because the model defaults to its highest reasoning effort, he wrote. He recommends turning that down for ordinary local use. Newer inference software may narrow the speed gap. Willison reported a roughly 72 percent performance gain on his Nvidia machine after switching on a feature the model supports, called Multi-Token Prediction.

The investor Tomasz Tunguz found a similar trade-off in a small test of the model in his own agent stack. With reasoning on, Qwen edged ahead on quality, he wrote, but it ran far slower than a cloud model. He cautioned that his nine tasks were too few for a verdict, and argued that the right measure is time to a finished answer, not raw token speed.

Why Alibaba is going small

The release fits a documented pattern. Hugging Face data reported by Business Insider shows that real usage skews toward smaller models, VentureBeat noted, even as huge frontier launches take the headlines. Models above 70 billion parameters made up only a small share of 2026 downloads.

Alibaba has leaned into that. It has published Qwen models across several size classes, and the strategy has already made the family the most downloaded open-weight line. The company says a managed cloud version of the model, with a one-million-token context window and built-in tools, is coming later. For now, the pitch is simpler. A file small enough to keep on a laptop can do work that recently belonged to the largest hosted systems.