Instinct

Instinct

Predicts the next edit you are likely to make in your code. A 7.6B model from Continue, fine-tuned from Qwen2.5-Coder-7B, for developers who want to run it on their own machine or server.

Made by continuedev on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Instinct is a next edit model from Continue. It reads the edits you are making in code and predicts the change you are likely to make next. Continue trained it further from the Qwen2.5-Coder-7B base using real code edits. The model has 7.6B parameters.

Who it suits: Developers who use Continue and want edit suggestions from a model they run themselves, on a laptop or on a shared server.

Industry | Who uses it | How they use it | Example
Software development | Developers who use Continue | Connect a self-hosted copy to the editor | Editing a function and getting the likely next change suggested
Individual programming | Developers with one machine | Run the quantized copy through Ollama | Starting the small local copy before a coding session
Platform teams | Teams that host their own models | Serve the model with SGLang or vLLM | Starting one server that several editors connect to

Skip it if you do not write code, or if you have no way to serve a 7.6B model.

Where to use it: It runs wherever you can serve a model and connect it to Continue.

• Continue: the editor tool the model is meant to be connected to
• Ollama: through a separate quantized copy for local use
• SGLang: a server you launch from the command line
• vLLM: another server option for the same weights
• Windows, macOS or Linux: where you can download the files with Python

Use case | Who | What they get back
Next edit suggestions | Developers in Continue | The change the model expects you to make next
Local inference | Individual developers | A smaller quantized copy that runs on your own computer
Self-hosted serving | Teams | A server that Continue can connect to

How to set it up: Decide first how you will serve it, then fetch or run it.

• Python with pip, to download the files
• Continue, to connect the model to your editor
• Room for the weights: about 15.2 GB in BF16, or about 3.8 GB at 4-bit as an estimate

1. Install the Hugging Face hub library.
2. Download the model files to your computer.
3. Serve them with SGLang or vLLM, or run the quantized copy with Ollama.
4. Connect the running model to Continue.

pip install -U huggingface_hub
hf download continuedev/instinct

To serve it with SGLang:

python3 -m sglang.launch_server --model-path continuedev/instinct --load-format safetensors

To serve it with vLLM:

vllm serve continuedev/instinct --served-model-name instinct --load-format safetensors

To run the quantized copy with Ollama:

ollama run nate/instinct

How to use it: Once a server or Ollama is running, point Continue at it and keep editing as usual. Each command below starts a different way of running the model.

Tool | What it does
hf download | Fetches the model files to your computer
ollama run | Runs the quantized copy locally
sglang.launch_server | Serves the full weights through SGLang
vllm serve | Serves the full weights through vLLM

Example: You type ollama run nate/instinct in a terminal. Ollama starts the quantized copy of Instinct on your machine. You then point Continue at it and edit some code, and it offers the change it expects you to make next.

For newcomers and experts: If you are new to this, start with the Ollama command. It is the shortest way to get the model running, and then you connect it to Continue and try a small edit.

If you know the field, this is a 7.6B model with Safetensors weights, built on Qwen2.5-Coder-7B and trained on real edit data. You can serve it with SGLang or vLLM, or use the Q4_K_M GGUF copy for lighter local inference.

You edit code in Continue
Continue calls the served model
Instinct predicts the next edit
You review the suggestion

Good to know: Instinct is a text generation model for the transformers library. Other people have published 8 quantized versions and 1 fine-tuned version of it.

• License: Apache 2.0
• Cost to run: none named, you supply the hardware
• Data it reaches: the card names none beyond the editor you connect it to
• Not for: the card lists no exclusions

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 7.6B
• Weights: about 15.2 GB in BF16; about 3.8 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 112
• Likes: 72
• License: apache-2.0
• Based on: Qwen/Qwen2.5-Coder-7B
• Quantized versions: 8
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen2, text-generation, conversational

Links

• Model: https://huggingface.co/continuedev/instinct
• Files: https://huggingface.co/continuedev/instinct/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download continuedev/instinct

Then follow the usage in its model card. Nothing in this kit runs by itself.