On March 11, 2026, the open-source Lemonade server shipped version 10.0 with Linux NPU support powered by the FastFlowLM runtime—the first working end-to-end path for LLM inference on AMD Ryzen AI NPUs under Linux. For over two years, AMD had been developing the AMDXDNA kernel driver, now merged into mainline Linux, to expose Ryzen AI NPU hardware to user-space software. Despite that infrastructure investment, practical tooling remained nearly absent: even AMD's own GAIA software fell back to Vulkan on the integrated GPU rather than using the NPU. Lemonade 10.0 and FastFlowLM 0.9.35, released simultaneously, close that gap.
FastFlowLM describes itself as an NPU-first runtime built exclusively for AMD Ryzen AI hardware. Version 0.9.35 adds official native Linux support and enables context lengths up to 256,000 tokens on current-generation Ryzen AI NPUs. Lemonade 10.0 layers on top of FastFlowLM to provide a server interface for running LLMs and Whisper speech recognition, along with a notable addition for the agent developer community: native integration with Anthropic's Claude Code agentic coding tool. Users must be running Linux kernel 7.0 or apply AMDXDNA driver backports to stable kernel versions to accommodate last-minute driver changes. Supported hardware covers all AMD Ryzen AI 300 and 400 series SoCs, including the Ryzen AI Max+ 395, the newly launched Ryzen AI Embedded P100, and the Ryzen AI PRO 400 series. Phoronix founder Michael Larabel, who broke the story, noted he plans to benchmark the NPU support on available lab hardware including Framework Desktop and Ryzen AI 300 Strix Point systems.
The timing matters because the Ryzen AI Embedded P100 and Ryzen AI PRO 400 are entering enterprise and embedded markets where Linux is the default—a very different audience than the consumer Windows segment that has dominated Ryzen AI's install base. For that audience, a working NPU inference stack is table stakes, not a bonus feature. AMD's competitors aren't close. Qualcomm's Snapdragon X Elite offers comparable NPU performance at 45 TOPS versus AMD's 50 TOPS, but remains effectively inaccessible for LLM inference on Linux: Qualcomm's primary NPU SDK is closed-source and Windows-centric, its proposed Linux kernel driver is still at RFC stage, and German Linux PC maker Tuxedo Computers publicly cancelled its Snapdragon X Elite laptop after 18 months, citing the chip as "less suitable for Linux than expected." Intel's OpenVINO framework provides the most mature documentation among x86 NPU vendors on Linux, but Intel Lunar Lake tops out at 48 TOPS. As of March 2026, AMD leads the Linux NPU inference space with shipping, open-source software that developers can run today.