Qwen 3.8 27B Ships Open Weights That Fit Into 17 GB

The accounting software vendor whose client financials are contractually barred from leaving its own servers has had the same AI plan for years: wait until open models are good enough. On Friday that wait got a little shorter.

Alibaba's Qwen lab released the weights for its new 27-billion-parameter model on August 16 under an Apache 2.0 license. You can download it, run it on your own hardware, ship it inside a commercial product, and pay nobody a licensing fee.

What's in the box

The model is 27B parameters, vision-capable, and handles a context window of 262,144 tokens. Quantized to Q4_K_M it occupies about 17 GB on disk, which is the number that moves it out of the data center and onto a well-specced workstation.

Simon Willison ran it on release day on a 128 GB MacBook Pro and on an NVIDIA DGX Spark, getting 15 to 30 tokens per second through LM Studio. Hosted APIs serving the same model run at 74 to 184 tokens per second. Turning on multi-token prediction speeds up the local figure by roughly 72%.

Do not judge it on the default setting

Willison's most practical finding is about configuration rather than capability. The model ships with xhigh reasoning effort enabled, which he calls "absolutely not a good way to run the model, especially on consumer hardware." It will think for minutes on a trivial question. His advice is to drop reasoning to low or turn it off for most work.

This matters because a fair number of teams try an open model, watch it crawl through a simple prompt, and conclude that open weights are not ready. Some of them are testing a bad default rather than a bad model.

How good is it?

Qwen claims the model beats both its predecessor, Qwen 3.6 27B, and the closed-weight Qwen 3.7-Plus. Willison found it competitive in some tasks with the much larger 2.4-trillion-parameter Qwen 3.8 Max released earlier this month. He rated it strong on complex coding work, reliable tool calling, and surprisingly precise bounding-box detection on images.

Part of that is the vendor's own measurement. What it does on your workload is a question only your data can answer.

Is it worth running yourself?

For companies operating under data-residency requirements, or ones that simply do not want to file cross-border transfer paperwork for customer records, a release like this is worth more than another round of price cuts from the closed providers. The hardest compliance question is rarely how capable a model is. It is where the data goes, and self-hosting mostly answers it.

Two caveats, honestly stated. Speed first: 15 to 30 tokens per second is marginal for a live chat interface. It is entirely adequate for overnight batch work such as document classification, extraction, and summarization. Hardware second: a 128 GB machine is not cheap, though it compares reasonably against a year of enterprise API spend.

Before you commit, build a small evaluation set of 50 to 100 examples drawn from your own documents and run it against both the open model and whatever you use today. Benchmark tables answer a general question; that test answers yours.

Sources: Simon Willison, Hugging Face / Qwen