The issue has been published 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 и Linux, and installation is available via the amd-gaia package. The v0.20.0 tag itself was published on June 3rd, but the release only appeared in the news feed on June 4th or 5th.
The main change in this version is the normalization of the execution device selection for each agent. Previously, GAIA used the GPU by default via the llama.cpp-based backend and did not provide a convenient way to switch a specific agent to the CPU or the energy-efficient Ryzen AI NPU. In GAIA 0.20.0, agents can declare supported devices, and the user selects the CPU, GPU, or NPU via the Agent UI or the CLI flag --device {cpu,gpu,npu}. GPU remains the default, and the gaia init --profile npu profile handles NPU detection, FLM backend installation, and model loading.
Changes in the release:
CPU/GPU/NPU selection for individual agents. Allows you to run demanding scenarios on the GPU, while less demanding or background scenarios run on the NPU or CPU. This is especially important for Ryzen AI owners: the NPU can be used for local inference with lower power consumption, without consuming the GPU.
Agent Hub TUI. Launching gaia without arguments now opens the agent management terminal. This allows you to view, search, launch, and manage agents without a graphical interface. The developers claim the standalone binary is approximately 21 MB in size and starts in less than 200 ms.
Stricter control of MCP tools. A second control level, activations, has been added for MCP connectors. Now, a connector assigned to an agent is no longer required to automatically add all its tools to the prompt: tools only become visible after being explicitly enabled for the connector-agent pair. This reduces prompt noise and helps small models more accurately select the required actions.
Accelerated mail processing. Email Agent has received seven new batch tools for mass inbox management. According to the developers, a typical workflow has been reduced from approximately 13 LLM requests to 2-3 steps, processing time from 488 seconds to 30-60 seconds, and token consumption from approximately 12 to 1,2.
RAG by PowerPoint files. GAIA now directly indexes .pptx files: text, tables, speaker notes, and embedded images via VLM analysis. Previously, users were prompted to first save presentations as PDFs.
Strengthening the safety and resilience of the first launch. This release closes a write-protection exploit via symbolic links in Python 3.10/3.11, extends write restrictions to four more file tools, and fixes a corrupted model diagnostic error that could trigger a re-download of approximately 25 GB of data.
GAIA is being developed as a local alternative to cloud-based AI services: data remains on the user's machine, and execution can be transferred between CPUs, GPUs, and NPUs depending on the task. For AMD, this is also a demonstration of the practical application of Ryzen AI not only as a marketing gimmick within the processor, but also as a dedicated computing device for local agents.
Source: linux.org.ru
