AMD GAIA v0.25.0 changes how its desktop app handles available memory, which models it selects for certain hardware, and what gets reprocessed after an upgrade. Phoronix placed the release in the week of October 19, 2026. AMD describes GAIA as agent software and a framework for building agents, rather than a language model. Its v0.25.0 release notes list the changes and installation options.
How GAIA adjusts to available memory
At first launch, the desktop app checks free memory and reduces the model’s context window as needed. The context window is the amount of text the model can consider at once; GAIA’s adjustment stops at 32K tokens. If memory is still insufficient, the app reports the shortfall and suggests closing applications.
A separate rule applies to PCs with less than 24 GB of RAM: GAIA loads the model with the first message. That threshold describes when model loading happens, while the 32K figure is the lowest context size the app will use during its memory adjustment.
Which models GAIA selects for specific hardware
AMD’s gaia init command chooses a model based on these stated configurations:
| Hardware configuration | Model selection | How it is selected |
| Strix Halo with at least 64 GB, or a GPU with 32 GB | Qwen3.6-35B-A3B | Selected by gaia init |
| Other configurations | Gemma-4-E4B | Default selection |
| Strix Halo with 128 GB | Qwen3.8 Flash Next | Optional; select it explicitly |
These are hardware-specific selection rules, not a blanket compatibility statement for every PC or model. The 128 GB Strix Halo option is opt-in rather than the default.
Changes to chat, speech, and Windows installation
The gaia chat command now runs AMD’s flagship GAIA agent instead of the retired ChatAgent. gaia talk uses Lemonade for transcription and speech, with Whisper and kokoro-v1; if Lemonade is unavailable, the command stops and reports the reason.
On Windows, AMD says the desktop app’s Python environment shrank from about 0.9 GB to about 0.1 GB after PyTorch, transformers, and accelerate were removed. That figure applies to the Python environment, not the size of the entire installer. The Windows installer can offer the desktop app, the terminal, or both.
The Agent UI also guides setup for the provider, model, and permissions, and its “always allow” grants can be revoked. Long turns continue while making progress; the default timeout is 75 minutes without progress, and users can change it through GAIA_STREAM_TIMEOUT.
What AMD reports about Qwen3.6 response time
AMD says caching the new-chat prompt changed the first “hello” response on Qwen3.6 from about 43 seconds to about 21 seconds. That is the specific response-time example in the release notes, not a general speed result for every system or workload.
What an upgrade reprocesses
On first start, GAIA re-embeds memories created with the retired nomic embedder. Embeddings are numerical representations used to retrieve relevant information. The software also re-indexes documents once; the first query after upgrading rebuilds the document cache.
The release changes permission handling as well. A skill that requests more permissions than its tier allows will no longer load, even if it is already installed on disk.
The Gemini issue AMD lists
AMD identifies a failure affecting Gemini models on AMD LLM Gateway when used with the Hub-installed GAIA agent. It occurs in the Agent UI and in the terminal interface’s second and later sessions. AMD advises choosing another gateway model and marks the fix as in review.
