A 16GB AMD graphics card ran a 27B Qwen model at about 36 tokens a second

What the test showed

As China.com reported, the outlet ran Qwen3.8-27B on a PowerColor Radeon RX 9070 XT Dark Dog with 16GB of video memory. The model has 27 billion parameters (the adjustable numbers that make up its knowledge) and can also interpret images. The name is printed as 'Qwen3.8' in the source, and we repeat it as written.

The software was LM Studio, a free desktop app on Windows. According to the report, it detects the AMD card on its own and picks the right acceleration, so setup is done by clicking rather than coding.

Choosing a file that fits

Open models usually come in quantized versions, meaning compressed copies that trade some quality for a smaller size. The outlet says a common 4-bit version of this model is about 16GB, which would fill the card and push work into slower system memory during long chats. A 2-bit version, around 8 to 9GB, leaves plenty of room but loses some reasoning and language quality.

It settled on a middle option called IQ3_XXS, at about 10.1GB. That left space for a roughly 931MB add-on that lets the model look at images, plus room for memory of the conversation.

Settings and results

The model supports up to 256K tokens of context (how much text it can keep in mind), but the outlet says maxing that out would nearly exhaust the card. It used 128K instead, with a compressed cache setting called q4_0 that needs about 2.2GB. Total use came to about 13.5 to 14GB, leaving 1.5 to 2GB spare. All model layers ran on the graphics card.

The test prompt asked for a snake game in HTML. The first output appeared after about 4.61 seconds, and generation held at 36.57 tokens per second. A token is a small chunk of text, roughly a word or part of one.

The model produced 19,344 tokens over about 533 seconds, finished normally and did not crash. The outlet says the resulting game worked, with responsive controls, scoring, speed-ups, wall collisions and sound effects.

Caveats

This is one outlet's single run on one card, with a coding prompt. The piece also promotes the PowerColor card's cooling design, so read it partly as hardware marketing. It gives no comparison with an Nvidia card or with other file sizes.