AMD GAIA 0.25.0 RC2 expands local AI and tightens agent safeguards

AMD’s GAIA 0.25.0 release candidate brings a larger local model, a much leaner Windows desktop install and tighter checks around agent permissions and files. It is a substantial preview for testers, not the finished release, and several changes may require users to rebuild data after upgrading.

AMD Watch analysis

What happened

AMD’s GAIA 0.25.0-rc2 release notes describe a release candidate with 300 commits since v0.24.1. GAIA now selects Qwen3.6-35B-A3B for systems with a 64 GB or larger Strix Halo, or a 32 GB GPU, and sizes the context window to available memory. Other systems retain Gemma-4-E4B by default.

The Agent UI now focuses on the GAIA agent, while the CLI and evaluation harness can still run other agents. The desktop app’s Windows Python environment drops from about 0.9 GB to 0.1 GB, according to the notes. The release also adds a single Windows installer for the desktop app and terminal.

Our top picks

• A larger local model on suitable hardware
  GAIA can select Qwen3.6-35B-A3B where the available GPU memory supports it.
• A much smaller Windows desktop environment
  Its stated Python environment falls from about 0.9 GB to 0.1 GB.
• Permission checks follow skills at load time
  A skill exceeding its permission tier is refused, even if it is already on disk.
• File tools check allowed folders
  Checks also cover paths inside inline shell scripts, extending the boundary beyond ordinary file operations.
• Declined actions stay declined
  GAIA does not retry a rejected action through another tool; it asks again with the exact command.
• Local Qwen tool calls get corrected sampling
  The notes say this fixes repeated tool-call loops by using the model card’s sampling settings.
• Long PDFs can index in the background
  The release also fixes document citations and ensures chunks are embedded whole, not just their first 1,200 characters.

Why it matters

This preview couples a more ambitious local-model option with practical changes to setup, permissions and document handling. For users with compatible hardware, model selection now responds to memory capacity; for everyone else, the default remains smaller. The slimmer desktop environment could make installation less of a storage commitment.

Some upgrades need care: memory saved with the retired Nomic embedder is re-embedded on first start, and indexed documents are re-indexed once. The new release also changes the Agent UI’s scope, so users who rely on its former agent picker should check the notes before moving over.

Our read

There is enough here to make this more than a version-number bump, especially for local-AI users and anyone who cares where an agent can write or what it can run. But it is still a release candidate. Test it if you want to help shake out the new model and safeguards; production users can wait for the final release, particularly if a one-time rebuild would be disruptive.

What to watch

• Whether the final release keeps the same hardware-based model selection.
• How the Agent UI changes land for users of custom agents.
• Whether the first-start memory and document rebuilds cause friction in practice.

Discussion spark: Would you trust a local agent more if it strictly enforced permission limits and refused to retry a declined action, even if that sometimes interrupts your workflow?

Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.