AMD is delivering Day 0 support for Qwen 3.8 27B, giving developers a path to run the dense open-weight model locally on AMD-powered PCs and workstations from the moment it becomes available. The company’s announcement, published August 14, targets systems with AMD Ryzen AI Max+ processors or a single AMD Radeon AI PRO R9700 32 GB graphics card, with support through open frameworks like llama.cpp.
Qwen 3.8 27B continues the Qwen family’s focus on coding, real-world work, research, and long-horizon AI workloads, and its 27 billion-parameter dense design places substantial demands on memory capacity and compute. AMD says the model can also run on supported AMD hardware with more than 24 GB of Variable Graphics Memory or VRAM. Early testing shows up to 24.5 tokens per second on a Ryzen AI Max+ 395 and up to 51.8 tokens per second on a single Radeon AI PRO R9700, preliminary figures measured on Windows using llama.cpp with the Vulkan backend across three or more runs. AMD expects performance to keep improving as Day 0 support matures.
LM Studio and Lemonade as entry points
For power users, LM Studio provides a graphical path to discover, download, and run the model on recommended AMD hardware, including Ryzen AI Max+ systems and the Radeon AI PRO R9700. AMD notes the model needs roughly 24 GB of VGM or VRAM to run comfortably, that MTP should be set to 4 on Ryzen AI Max+ and 2 on the Radeon AI PRO R9700 for optimal performance, and that the “Try mmap” option should be unchecked in advanced model load settings.
The headline figures come with a specific test setup. AMD measured the numbers on Windows 11 using llama.cpp with the Vulkan backend, running a GMKtec EVO X2 AI Mini PC with a Ryzen AI Max+ 395, 128 GB of system memory, and VGM set to 64 GB, with throughput averaged across three or more runs. The company flags the results as preliminary, with additional software and model optimizations still underway, which is why the 24.5 and 51.8 tokens-per-second figures are best read as an early floor rather than a ceiling.
For application developers, AMD points to Lemonade, its local-first developer platform designed to reduce the complexity of deploying AI across PCs. Lemonade’s multi-engine architecture provides a unified interface while handling hardware-aware backend selection and optimization across available CPU, GPU, and NPU resources. It can be packaged alongside an app as a lightweight local inference layer, letting applications talk to Qwen 3.8 27B through familiar API patterns rather than a hardware-specific inference stack.
Day 0 support matters because open models ship on a fast cadence, and AMD says users should not have to wait for their hardware and software ecosystem to catch up each time a new model arrives. The release extends the case for local AI on hardware you control, and a 27B model running fully on-device is exactly the kind of workload where memory bandwidth decides how usable the result is. It also lands alongside AMD’s broader push to make on-device inference a first-class platform feature, the same week Google detailed the on-device AI capabilities of its Tensor G6 chip. The model is ready in LM Studio today on supported AMD systems, with Day 0 support marking the start of what AMD says will be an ongoing optimization effort.
