AMD GAIA 0.25 Cuts Its Local AI Agent Install From 0.9 GB to 0.1 GB
GAIA is AMD's open-source tool for running local AI agents on Ryzen AI PCs. Version 0.25, released 18 October, removes PyTorch, transformers and accelerate so the Windows Python environment drops from about 0.9 GB to 0.1 GB, checks free memory before loading a model so 16 GB machines no longer crash on first launch, and defaults to Qwen3.6-35B-A3B on 64 GB Strix Halo systems.
On this page
A much smaller install
GAIA began life as "Generative AI Is Awesome" and is now described by AMD as its open-source agent factory: a tool for building and running local AI agents on Ryzen AI hardware, on Windows and Linux, with specific optimisations for Strix Halo. The repository is MIT licensed and pitches itself simply as "Build AI agents for your PC."
The headline change in 0.25, tagged on 18 October, is weight loss. The release notes state that "PyTorch, transformers and accelerate are gone; the desktop app's Python environment drops from ~0.9 GB to ~0.1 GB on Windows." Windows users also get a single installer that covers the desktop app and the terminal, and adds a Windows Terminal profile. Dropping the training-oriented libraries from an inference tool is the kind of cleanup that rarely makes headlines, but a tenfold reduction in the environment that ships to every user is the difference between a quick download and an abandoned one.
The 16 GB fix: shrink the context, not the model
The second change addresses a failure that would have hit exactly the users AMD most wants to win. On a 16 GB PC, the desktop app's first launch could run out of memory and get killed by the Linux kernel's OOM handler. The fix is pragmatic: "GAIA checks free memory before it loads a model, shrinks the context window to fit (to no less than 32K)." PCs with under 24 GB of RAM now load the model on the first message rather than at startup.
That approach deserves a note for anyone building local AI software. The instinct when memory is tight is to reach for a smaller model. GAIA instead keeps the model and trims the context budget, which preserves answer quality while trading away long-conversation headroom. It is the same trade-off that governs responsive chat on phones, where every prompt token costs prefill time and memory, and where a shorter context is usually the right default unless the user explicitly asks for more.
Bigger defaults on Strix Halo, and a faster first reply
At the other end of the hardware range, 0.25 gets more ambitious. The gaia init command now "picks Qwen3.6-35B-A3B on a 64 GB+ Strix Halo or a 32 GB GPU and keeps Gemma-4-E4B elsewhere," with Qwen3.8 Flash Next available as an opt-in on 128 GB Strix Halo machines. Prompt priming cuts the time to a first "hello" on Qwen3.6 from roughly 43 seconds to roughly 21 seconds. Document and memory indexing has moved to the GPU, which the notes say frees about 11 CPU cores on Strix Halo.
Long agent turns are handled differently too. Rather than a fixed clock, GAIA now times out only after no progress is made, with a 75 minute default that can be changed through the GAIA_STREAM_TIMEOUT variable. For users who mix in hosted models, DeepSeek V4.1 Flash becomes the default cloud model on the AMD LLM Gateway and Fireworks.
What breaks for existing users
This is a feature release with real breaking changes. The gaia chat command now runs the flagship GAIA agent, the same one used by the TUI and the Agent UI, instead of the retired ChatAgent. The gaia talk command transcribes and speaks through AMD's Lemonade local server using Whisper and kokoro-v1, and stops with a reason if Lemonade is down; the separate audio extra is removed.
The Agent UI now runs only the GAIA agent, so the Agent Hub home, the agent picker and custom-agent import and export are gone from the interface. The retired nomic embedder is removed, which means memory saved with it is re-embedded on first start, and indexed documents are re-indexed once on the first query after upgrading. A skill that asks for more than its tier allows no longer loads, even if it is already on disk. One known issue remains: Gemini models on the AMD LLM Gateway fail with the Hub-installed agent, with a fix in review.
For readers who already run local models through CuriousLM or similar tools, GAIA is worth watching less as a competitor than as a signal. A chip vendor shipping an MIT-licensed agent runtime, choosing a 35B mixture-of-experts model as a default on consumer hardware, and solving memory pressure by budgeting context rather than downgrading the model, is a vendor that expects local AI to be ordinary.