Copilot with a local model: what GitHub actually allows and what isn't available offline yet

For developers at banks, in the healthcare sector, or at companies with strict security policies, the question is simple: Is it possible to use an AI assistant for programming without sending the code to the cloud? Reports say that GitHub Copilot has made this possible since October 23, 2026, through a public beta feature called “Copilot Local.” However, we haven’t found such an announcement. The reality is more complex: Copilot has supported its own local models for some time, but “local” does not automatically mean “offline” or “the code never leaves the computer.”

What is confirmed

According to official announcements, GitHub is gradually expanding support for custom models.

• On April 7, 2026, the Copilot CLI tool gained support for custom model providers as well as local models, such as those via Ollama, vLLM, or Foundry Local. When using a custom provider, logging into GitHub is not required.
• On June 23, 2026, support for agent sessions was added to the Copilot app, according to one analysis.
• On October 7, 2026, starting with version 1.0.94, the Copilot CLI began displaying models running locally via Ollama in the model list.

In VS Code, custom models are added by configuring custom endpoints and using the language model editor. Ollama also offers an extension that adds its models to the model selection in the VS Code chat. According to Copilot’s description, this page does not mention it by name.

What the “local model” does not mean

It isn't automatically offline

GitHub explicitly states that selecting the local model does not enable offline mode or disable telemetry. For the Copilot CLI, there is a separate offline mode that is enabled via the variable COPILOT_OFFLINE=true. In this mode, the CLI does not communicate with GitHub’s servers, telemetry is disabled, and it communicates only with the provider you have configured. The combination of the local model and offline mode allows you to work even in environments without internet access.

You also need to set up helper models in VS Code

According to one description, VS Code also requires selecting a fast helper model for internal tasks, such as conversation names or commit messages. If you do not set it to local, these tasks may use the default hosted option. For companies with a policy prohibiting the transmission of data, this is an important detail.

What We Don't Know

We found no confirmation in our sources that the local model is also used for automatic code completion while typing (so-called “ghost text”). Nor did we find any official settings with the names mentioned in some guides, such as github.copilot.advanced with the following items localEndpoint or offlineMode. We therefore do not recommend copying them. Likewise, we have not confirmed the specific model “llama3-coder:8b” or extension version v1.250.

Documented procedure for Copilot CLI

Both GitHub and Ollama document this principle for their terminal versions: environment variables are used to configure the local server and model, and optionally, offline mode.

export COPILOT_PROVIDER_BASE_URL=http://localhost:11434
export COPILOT_MODEL=NAZOV-VASHO-MODELU
export COPILOT_OFFLINE=true
copilot

The model must support tool calls and streaming. Before deployment, verify the exact variable names in the current GitHub documentation, as they are subject to change.

What hardware is required?

We did not find any specific requirements from the manuals, such as “8 GB VRAM” or “16 GB of memory,” in the official materials. Generally speaking, small quantized models run on a standard developer computer, but larger models require significantly more memory and may be slower to respond. This is a general observation, not verified data from a single source. Therefore, before deploying the model in your company, we recommend testing the specific model on the specific hardware.

Compromises Compared to the Cloud

• Quality. Small local models typically do not handle complex architectural tasks as well as the largest hosted models. They may be sufficient for routine updates, test generation, or documentation, but it depends on the model and the task.
• Features. Agent-based features require a model with tool support, which not every local model provides.
• Cost. Starting June 1, 2026, GitHub will charge for Copilot on a pay-as-you-go basis. Local models eliminate model usage fees but require dedicated hardware.

What does this mean for businesses?

For strictly regulated environments, it is essential that all components be verified: the model itself, offline mode, auxiliary models, telemetry, and the tool’s network communication. If the code is to remain on the device, we recommend monitoring network communication—for example, through a firewall—and verifying the procedure with the security department.

What we haven't confirmed

• The public beta of “Copilot Local” from October 23, 2026. We couldn’t find any mention of it in the official announcements.
• Fully offline operation in VS Code “without internet” after a single setup. Sources point out additional settings and limitations.
• Hardware requirements and a speed of 30 tokens per second. The data comes from unverified community tests.

Conclusion

GitHub Copilot actually runs on its own local model, most advanced in the terminal, where there is also a separate offline mode. However, headlines about a simple “Copilot without the internet” in VS Code oversimplify the reality. Companies with strict policies should verify the process in the official documentation and in practice, not just by relying on copied settings.