What changed in the picker
GitHub’s 7 October CLI update adds local Ollama model discovery through /model, starting with version 1.0.94-0. Ollama and the model must already be installed and running; discovery does not download them. A discovered model is added only after confirmation, with a choice to switch for the current session or add it without switching.
Models need tool calling and streaming support. The announcement says choosing a local model does not activate offline mode or disable GitHub telemetry.
What offline mode actually changes
GitHub’s provider documentation describes COPILOT_OFFLINE=true as an explicit setting to prevent contact with GitHub’s servers. If the configured provider endpoint is remote, prompts and code context can still go to that provider. Offline mode therefore depends on where the provider actually runs.
The docs support OpenAI-compatible, Azure and Anthropic provider types, including local runtimes. Supplying a provider yourself also means checking that provider’s setup and access requirements. Discovery is an easier selection flow, not evidence of the selected model’s quality.
Four fields for a useful local-model receipt
Before comparing models, write down four independent facts about the session. This proposed record keeps a friendly model name from standing in for the whole configuration:
- Model identity: the exact model identifier and runtime version, with the capabilities required by the client. Record what is selected for this session, not just what appears in the picker.
- Endpoint: the host that receives inference requests. Keep secret tokens out of the record. A familiar model name is insufficient to establish this location.
- Network behavior: the explicit mode chosen and, if you have suitable monitoring, the observed destinations during a small synthetic request. Record the observation window and what that monitoring cannot see.
- Tool access: which files, commands and other integrations the session can use. Treat those permissions as a separate part of the setup.
For a comparison task, use a disposable repository with a failing test whose expected result you can explain. Preserve the initial commit, task text, produced diff, actual test output and elapsed time. Count retries and manual edits too. One completed request cannot establish general coding accuracy; a repeatable fixture makes its narrow result inspectable.
What remains untested here
We checked the official announcement and documentation. We did not install a runtime, download a model, capture network traffic or run a native Copilot task. No local model receives a Fewertools ranking from this update. See the Copilot profile and the separate local sandboxing report for execution-policy questions.
Sources and scope
Official sources read 10 October 2026. Facts above are attributed to the provider; the evaluation worksheets are proposed checks. This article does not report a native product test or assign a tool score.