AMD has released Day 0 support for the dense Qwen3.8 27B artificial intelligence model across its PC and workstation hardware. The rollout allows developers to run the heavy 27B parameter model locally through open tools like llama.cpp and LM Studio, hitting speeds up to 51.8 tokens per second. But taking advantage of this setup requires a capable machine packing at least 24 GB of graphics memory.
Dense models demand serious memory bandwidth. AMD shared early testing data showing how the model performs on modern silicon using the Vulkan backend in llama.cpp. A dedicated workstation equipped with a single AMD Radeon AI PRO R9700 graphics card and 32 GB of memory reached 51.8 tokens per second. On the mobile side, an AMD Ryzen AI Max+ 395 processor pushed 24.5 tokens per second. Both configurations provide enough speed for real time coding and research tasks.
Running this architecture requires plenty of hardware headroom. The 27B model needs roughly 24 GB of dedicated VRAM or Variable Graphics Memory to operate without crashing. AMD notes that optimizations are still ongoing, meaning token throughput should climb as software drivers mature.
Getting the model running takes minimal setup for creators who want a visual interface. LM Studio now supports Qwen3.8 27B directly on compatible AMD hardware, allowing users to download weights and test prompts without touching code. Users simply need to adjust their draft token settings to get maximum performance out of the engine.
For software creators who want to build custom tools, AMD is pushing its Lemonade developer platform. Lemonade serves as a lightweight local inference layer bundled directly inside applications. It automatically routes computational tasks across CPU, GPU, and NPU resources. This eliminates the headache of writing custom hardware drivers for every PC configuration, letting developers ship software that talks straight to the model.
