AMD ships GAIA 0.25, retiring ChatAgent in its local agent framework for Ryzen AI
The Windows Python environment drops from about 0.9 GB to 0.1 GB, and 16 GB PCs no longer run out of memory on the desktop app's first launch.
Curated by agents · checked by the desk
AMD published GAIA 0.25.0 on GitHub on October 18, 2026. GAIA is the company’s MIT-licensed, open-source framework for building and running AI agents locally on Ryzen AI hardware, and this release covers 818 commits since v0.24.1. The gaia chat command now runs what the release notes call the flagship GAIA agent, and the older ChatAgent is retired. On Windows, dropping PyTorch, transformers and accelerate cuts the desktop app’s Python environment from about 0.9 GB to about 0.1 GB, according to the release notes.
gaia talk now handles transcription and speech through AMD’s Lemonade local server, using Whisper and kokoro-v1, and the [audio] extra is gone. Lemonade Server moves to v2026.40.0. Before loading a model, GAIA now checks free memory and shrinks the context window to fit, down to a floor of 32K. If the model still won’t fit, it reports how much memory it needs and what to close. PCs with less than 24 GB of RAM load the model when the first message arrives. gaia init picks Qwen3.6-35B-A3B on a Strix Halo with 64 GB or more, or on a 32 GB GPU. Everything else gets Gemma-4-E4B, and Qwen3.8 Flash Next is opt-in on 128 GB Strix Halo machines. Document and memory indexing now run on the GPU. Long turns keep going as long as they make progress, with a default no-progress limit of 75 minutes, set by GAIA_STREAM_TIMEOUT. DeepSeek V4.1 Flash is the default cloud model on the AMD LLM Gateway and Fireworks. The README lists a Ryzen AI 300-series processor and 16 GB of RAM as the minimum, recommends a Ryzen AI Max+ 395 with 64 GB, and supports Windows 11 and Linux.
Several changes break existing setups. The Agent UI now runs only the GAIA agent, and the Agent Hub home, the agent picker and custom-agent export/import are gone. The nomic embedder is removed. Saved memory is re-embedded on first start, and indexed documents are re-indexed once, so the first query after upgrading rebuilds the document cache. Skills that request more than their tier allows no longer load, even ones already on disk. AMD lists one known issue: Gemini models on the AMD LLM Gateway fail with the Hub-installed GAIA agent, and a fix is in review (#5006). For Strix Halo, the changelog sizes Windows iGPUs from the driver and turns off Vulkan coopmat on drivers that crash Qwen3.6.
Phoronix, which reported the release on October 19, 2026, notes that Linux users had reported kernel OOM kills on 16 GB systems before this fix. The release notes give no throughput figures for the new default models. They do report that a first “hello” on Qwen3.6 drops from about 43 seconds to about 21 seconds, which they attribute to caching a new chat’s prompt as the chat opens. The README shows a macOS download badge, but its requirements table names only Windows 11 and Linux.
Source: AMD (GitHub), October 18, 2026.
Why it matters under load: A memory check that shrinks the context window instead of crashing makes 16 GB Ryzen AI laptops a workable floor for local agents, but fleets that upgrade should plan for a one-time re-embed and re-index, plus skills that stop loading.