Released version AMD GAIA 0.20.0 — an open framework for running local AI agents on PCs with AMD Ryzen AI hardware acceleration. The project is distributed under a license MIT, supports Windows and Linux, and installation is available via the amd-gaia package. The tag v0.20.0 was released on June 3, but it appeared in the news feed on June 4–5.
The main change in this version is the proper selection of execution devices for each agent. Previously, GAIA defaulted to using GPU via a backend based on llama.cpp and did not provide an easy way to switch a specific agent to CPU or energy-efficient Ryzen AI NPU. In GAIA 0.20.0, agents can declare supported devices, and the user can choose CPU, GPU, or NPU through the Agent UI or CLI flag —device {cpu,gpu,npu}. GPU remains the default option, while the profile gaia init —profile npu takes care of NPU detection, FLM-backend installation, and model loading.
Release changes:
Selection of CPU/GPU/NPU for individual agents. Allows heavy scenarios to run on GPU while less demanding or background tasks can run on NPU or CPU. This is especially important for Ryzen AI owners: NPU can be used for local output with lower power consumption, without occupying the video core.
Agent Hub TUI. When launching gaia without arguments, a terminal agent management center now opens. From there, you can view, search, launch, and manage agents without a graphical interface. Developers indicate the size of the standalone binary to be around 21 MB and launch time to be less than 200 ms.
Stricter control of MCP tools. For MCP connectors, a second level of control — activations — has been added. Now the connector assigned to an agent is not required to automatically add all its tools to the prompt: tools become visible only after being explicitly enabled for the "connector — agent" pair. This reduces noise in the prompt and helps smaller models to choose the necessary actions more accurately.
Accelerated email processing. The email agent has received seven new batch tools for mass organization of incoming messages. According to developers, the typical scenario has reduced from about 13 LLM requests to 2–3 steps, processing time from 488 seconds to 30–60 seconds, and token usage from about 12,000 to 1,200.
RAG for PowerPoint files. GAIA now directly indexes .pptx: text, tables, speaker notes, and embedded images through VLM analysis. Previously, users were required to save presentations in PDF first.
Enhancing the security and resilience of the initial launch. The release closes a bypass of write protection via symbolic links in Python 3.10/3.11, extends write limitations on four additional file tools, and fixes a corrupted model diagnostic error that could trigger a re-download of approximately 25 GB of data.
GAIA is evolving as a local alternative to cloud AI services: data remains on the user's machine, and processing can shift between CPU, GPU, and NPU depending on the task. For AMD, it also showcases practical applications of Ryzen AI not just as a marketing block in the processor, but as a standalone computing device for local agents.
Source: linux.org.ru
