Custom code completion comes to Copilot Enterprise
GitHub is opening a limited public beta for fine-tuned models in Copilot Enterprise, letting organizations adapt in-line code completion to their own repositories and engineering conventions. The move extends Copilot's existing customization features—repository indexing and knowledge bases, both built on retrieval-augmented generation (RAG)—into the real-time code completion path for the first time.
RAG is effective for chat scenarios where pulling current context from outside the model's training data matters. But it doesn't meet the latency demands of inline suggestions. Fine-tuning changes that: the custom model is trained on private codebases plus telemetry from how a team accepts or rejects Copilot's suggestions, so completions reflect internal modules, proprietary or legacy languages, and established coding style without the retrieval overhead.
Where fine-tuning fits
Teams that depend on internal APIs, specialized frameworks, or strict compliance and security standards stand to gain the most. For example, financial organizations working with legacy languages like COBOL can train the model against those codebases. Similarly, technology and healthcare companies that enforce internal-library usage to keep cloud deployments aligned with policy can see more accurate suggestions with less rework.
For developers, fewer corrections mean faster onboarding and more time spent building rather than fixing. The quality checks baked into the training pipeline are designed to ensure custom models outperform the baseline before they ever reach a developer's editor.
How custom models are built
Each custom model starts from GitHub's base in-line completion model and is fine-tuned using LoRA (Low-Rank Adaptation). LoRA tunes a small subset of the most significant model parameters during supervised training, which keeps the process faster and more cost-effective than traditional full-model fine-tuning. When base models are upgraded, subsequent retrainings roll forward to the newer base automatically.
The training pipeline runs on the Azure OpenAI Service, which gives the process the required scale and security boundaries. When a training run starts, repository and telemetry data are tokenized and temporarily copied into the Azure pipeline. A portion is used for training; the rest is held back for validation and quality assessment.
If the candidate model passes evaluation—including tests against your validation data to confirm improved suggestions for your repositories—it is deployed to Azure OpenAI. There, multiple LoRA models can be hosted at scale while remaining network-isolated from one another. Once the process finishes, temporary training data is removed from all surfaces, and inference traffic resumes through the normal Copilot proxy path, which routes each developer to the correct custom model.
Privacy and security guardrails
Two guarantees anchor the security posture: your data is never used to train another customer's model, and your custom model remains private and under your control.
Beta availability
The limited public beta opens today with capacity expanding gradually. GitHub says it will share updates as early adopters provide feedback and access broadens over time.
