A 27-billion-parameter AI model you can run at home just matched cloud-only rivals on coding tests

Key points

• Alibaba released Qwen3.8-27B on Hugging Face on Friday under an open Apache 2.0 licence, meaning anyone can download and modify it for free.
• Independent benchmarking firm Artificial Analysis scored it 52 on its Intelligence Index, equal to OpenAI's GPT-5.6 Luna cloud model at maximum reasoning.
• Compressed to roughly 17 gigabytes, the model runs on a high-end gaming PC or a recent Apple MacBook Pro with an M5 Max chip.
• The model passed 3 million downloads on Hugging Face within its first three days of release.
• Its default "think hard" setting can slow responses dramatically: one simple image request took 21 minutes before producing an answer.

Friday's most-talked-about AI release was not from OpenAI, Anthropic or Google. It came from Alibaba, the Chinese e-commerce and cloud giant, and it landed quietly on Hugging Face, a website where researchers share free AI models anyone can download.

The model is called Qwen3.8-27B. The "27B" refers to 27 billion parameters, which are the internal numerical settings that determine how a model thinks and responds. Bigger numbers generally mean more capability, but also more hardware needed to run the thing.

This one sits in a sweet spot.

What can it actually do?

Qwen3.8-27B handles text, images and video in a single model. It can write and debug code, answer questions about a photo, and carry out multi-step tasks on its own, the kind of thing researchers call "agentic" work, where the model plans and executes a sequence of actions rather than simply answering one question.

It also carries a long memory. Its context window, the amount of text it can hold in mind at once, stretches to 262,144 tokens. One token is roughly three-quarters of a word, so that is around 200,000 words of conversation or document it can process in a single session.

How does it compare to the big paid models?

Independent testing puts it squarely alongside models that cost money to use via the cloud. Artificial Analysis, a firm that benchmarks AI models using standardised tests, gave Qwen3.8-27B a score of 52 on its Intelligence Index, a composite of nine tests covering coding, science, reasoning and professional tasks. That matches the score it currently gives OpenAI's GPT-5.6 Luna at its most capable setting, a model available only as a paid cloud service.

On a separate index measuring agentic tasks, the model scored 51, beating Anthropic's Claude Opus 4.8, which launched less than three months ago.

Alibaba's own launch figures showed 61.7 on SWE-bench Pro (a standard software-engineering test), 90.3 on LiveCodeBench v6 (a coding challenge), and 84.3 on OSWorld-Verified (a test for navigating computer interfaces). Some of those comparisons use internal evaluation setups that do not match the exact conditions other labs use, so treat them as directional rather than definitive.

VentureBeat, which first reported the wider developer reaction, noted that Alibaba's benchmarks showed the 27B model beating Claude Opus 4.6 Max on SWE-bench Pro, though Opus still leads on several other tests.

What hardware do you need to run it?

That depends on how compressed the file is. At full quality, the model needs about 56 gigabytes of GPU memory, which is specialist territory. Compressed to a format called FP8, it needs roughly 28 gigabytes. Compressed further to a format called 4-bit quantisation, a technique that shrinks the file by rounding numbers aggressively with only a small quality penalty, it drops to around 17 gigabytes.

That 17-gigabyte version runs on a powerful gaming desktop or a MacBook Pro with an M5 Max chip. Developer Simon Willison tested exactly that configuration and found the model could write code, interpret images and run an autonomous coding loop entirely on his own machines without sending any data to an outside server.

"The fact that a 17GB file can do all of this stuff on my home machines is a miracle," Willison wrote.

Version	Memory needed	Typical hardware
Full 16-bit	~56 GB GPU	High-end workstation
FP8 compressed	~28 GB GPU	Professional GPU server
4-bit quantised	~17 GB	Gaming PC or M5 Max MacBook Pro

Is there a catch?

Yes. Qwen3.8-27B defaults to a mode where it "thinks hard" before answering, generating a long chain of internal reasoning before producing a response. That costs time and computing effort. Willison asked the model to draw a pelican riding a bicycle as an SVG image. It took 21 minutes and generated more than 22,000 internal reasoning tokens before producing the picture.

Investor and developer Tomasz Tunguz ran a small nine-task comparison against a rival model and found Qwen slightly ahead on quality but roughly 30 times slower and 4.5 times more expensive to run when reasoning was switched on. Both Willison and Tunguz recommend starting with reasoning set to low or off for everyday use.

Speed also varies by setup. Willison saw 15 to 30 tokens per second, which is noticeably slower than most hosted models. Enabling a feature called Multi-Token Prediction through a free tool called llama.cpp gave him a 72% speed boost on his machine, though hosted services still feel snappier.

What does this mean for businesses and ordinary users?

For companies, the calculation is straightforward. A model that runs inside your own building cannot leak data to an outside server. Apache 2.0 licensing means an organisation can inspect the code, modify the model and host it behind its own security controls, with no subscription fees or usage costs per query. Alibaba has confirmed the model works with popular serving frameworks including vLLM and SGLang.

For curious individuals, a free download and a capable home machine are all that stand between you and a model that, on independent tests, performs on par with services that cost money per use just a few months ago.

Download figures suggest people noticed fast. The model passed 3 million Hugging Face downloads in its first three days.