Instinct: A Next-Generation Code Editing Model

Instinct is a 7.6 billion parameter open-source model developed by Continue, specifically engineered for advanced code editing. Fine-tuned from Qwen2.5-Coder-7B, Instinct leverages a unique dataset of real-world code edits to intelligently predict and suggest the next logical code modifications, aiming to keep developers in a state of flow.

Key Capabilities

Intelligent Code Prediction: Predicts subsequent code edits based on real-world developer patterns.
Flow State Enhancement: Designed to integrate seamlessly into development workflows, minimizing interruptions.
Robust Fine-tuning: Built upon Qwen2.5-Coder-7B and extensively fine-tuned on a proprietary dataset of code edits.
High Context Length: Supports a substantial context window of 131072 tokens, allowing for comprehensive understanding of larger codebases.

Deployment Options

Instinct is available for various deployment methods, including:

Ollama: A Q4_K_M GGUF quantization is provided for efficient local inference.
SGLang & vLLM: Supports serving via SGLang and vLLM for integration with Continue's self-hosting options.

For more in-depth information, refer to the Continue blog post on Instinct.