Alibaba's Qwen3.8-27B Tops 7.7 Million Hugging Face Downloads

Alibaba's Qwen3.8-27B has passed 7.7 million downloads on Hugging Face, according to Hugging Face's own listing, a month after its Aug. 17 release.

Alibaba released the 27-billion-parameter model under an Apache 2.0 license alongside its flagship Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters, Alibaba Cloud said in its own announcement. It was the first time Alibaba released the weights of a Qwen-Max-class model openly, the company said.

The 27B model is dense rather than mixture-of-experts, natively multimodal across text, image and video input, and carries a native context window of 262,144 tokens that extends to 1 million, according to its Hugging Face model card. Alibaba said the full model needs 56 gigabytes of GPU memory, 28 gigabytes at 8-bit precision, or 17 gigabytes at 4-bit quantization, low enough to run on a single high-end consumer or workstation GPU.

What the benchmarks show

Benchmark	Qwen3.8-27B	Comparison
SWE-bench Pro	61.7%	53.4% (Claude Opus 4.6 Max)
GPQA Diamond	89.2%	not disclosed
LiveCodeBench v6	90.3%	not disclosed

The figures above are Alibaba's own. The company said the smaller model matches the output of Qwen3.7-Plus, a mixture-of-experts model roughly ten times its size, on real-world coding and office tasks, a claim The Terminal could not independently verify.

An independent reviewer pushes back

Kingy AI, a hardware-focused reviewer, said Qwen3.8-27B "reaches frontier-level performance only on select self-reported benchmarks, falling short of a true apples-to-apples win," and that the best existing systems still hold meaningful leads on tasks Alibaba left out of its own comparisons. Artificial Analysis, an independent benchmark provider, had not published its own test of the model as of Sept. 15. Simon Willison, a software engineer who writes about AI tools, called it "a miracle" that a model this capable compresses to a 17-gigabyte download that fits on ordinary hardware, according to Cybernews.

The release follows two other open-weight launches this summer: Meta's Muse Glimmer and Nvidia's Nemotron 3.5 Lightning, according to AI Business, a trade publication that covers the enterprise AI market. Bradley Shimmin, an analyst at Futurum Group, told AI Business that open releases let companies apply cost controls to AI spending the way they already do with other infrastructure, rather than depending on a single vendor's pricing.

The download count puts Qwen3.8-27B ahead of other recent open releases, including m-a-p's YuE2-3B music model and DeepSeek's V4.1-Flash, a mixture-of-experts model that needs far more GPU memory to run despite activating a similar share of its total parameters.

Derivative versions with safety guardrails removed are already circulating. One such variant, labeled "uncensored" by its publisher, has drawn more than 875,000 downloads of its own on Hugging Face, separate from the base model's count, which The Terminal confirmed against Hugging Face's public listings. Alibaba has not commented on the derivative releases.

Alibaba has not said whether Qwen3.8-Max's full weights will get the same consumer-hardware quantized releases the 27B model already has.