本機推論／開發工具
Ollama 0.34 RC Connects Local Models to ChatGPT Desktop, Adds Compatibility for Tool Search and Response Compaction
Ollama’s macOS app can now connect local open models directly to the latest ChatGPT Desktop. It remains a release candidate, and the community has reported an interface issue when switching back to cloud models, making it unsuitable for immediate large-scale deployment.
Ollama released version 0.34.0-rc1 on September 5. Rather than adding new models, its main change allows users to complete setup in the Ollama macOS app and then select Ollama-managed local models directly in the latest ChatGPT Desktop. This enables ChatGPT’s Chat, Work, and Codex desktop workspaces to integrate more directly with a local inference backend. Model weights and inference can remain on the user’s computer, eliminating the need to call a cloud model API for every generation.
The integration still requires protocol-level compatibility. This RC also adds support for tool search and response compaction in OpenAI-compatible clients, and fixes image delivery in compacted responses. For long-running agents, response compaction preserves necessary state as the context grows. If the local server fails to correctly reconstruct tools, images, or message identifiers, the agent may lose observations or even execute tools repeatedly. The new version also improves structured output performance on Apple Silicon, which is more important for workflows that rely on JSON Schema or tool-parameter validation than for ordinary chat.
However, a “local model” does not mean the entire workflow is offline. Whether ChatGPT Desktop’s cloud synchronization, search, external tools, and account features transmit data still depends on the active mode and permissions. Engineering teams should track the data paths for model inference, tool calls, and conversation storage separately. The official release page also labels version 0.34.0 as a prerelease, and the documented setup entry point is currently limited to macOS.
In addition, a single user in the OpenAI developer community has reported that after enabling Ollama, the desktop app displayed only local models and made the cloud options unavailable. This is not enough to establish a widespread regression, but it is a clear reason to test switching, sign-out, and recovery procedures. The next points to watch are whether the stable release expands support to Windows, whether connection and permissions documentation is published, and whether the desktop client can reliably switch between local and cloud execution.