So 🔥
I love when startups do research and just ship in product. Skip the benchmarking theatre.

This work was possible thanks to a generous compute grant from the Nvidia DGX Cloud Innovation Lab, allowing us access to a dedicated 8xH100 GPU instance for distributed training and inference

Instinct was the product of an incredible summer internship project by Adarsh, with guidance from the rest of the Continue team

We're just getting started with this first release and are particularly excited to push forward Next Edit by:
- post-training with KTO
- scaling the dataset by 10x

Instinct's weights, dataset, and training code are all available for use under the Apache-2.0 license

If you are interested in furthering the state of the art, either as part of the community or the Continue team, please reach out!

huggingface.co/continuedev/...

If you want to try the model, you can run it with @ollama and the Continue VS Code extension by following our guide

docs.continue.dev/guides/insti...

To evaluate Instinct, we adopted an LLM-as-judge strategy, similar to Zeta from Zed, asking Claude to rank edits on a scale from 1-5 based on functional equivalence to the ground truth edit

Instinct’s average score of 3.877 outperforms Zeta’s score of 3.735

Given that Qwen2.5-Coder-7B is already strong base model, we wanted to preserve its existing coding knowledge

We used Selective Knowledge Transfer (SeleKT), which computes dense gradients, but applies weight updates for only the top-k gradients

www.microsoft.com/en-us/resear...

Our dataset of 4,000+ high-quality real-world edits was built by collecting keystroke-level data from the Continue team and chunking it into coherent edits

For multilingual support we used Qwen3-Coder-30B to "translate" these into synthetic data for Java, C++, Python, and Rust

While traditional autocomplete inserts text by using a fill-in-the-middle prompt format (FIM), Next Edit accepts a prefix, suffix, and "range to replace", allowing it to make complex changes all at once by rewriting the range

To users, this appears as an inline diff

Our goal in training and open-sourcing Instinct is to lay a foundation for even stronger open models and enable self-hosting. Check out our blog for full details:

blog.continue.dev/instinct

Introducing Instinct 🐺, a new state-of-the-art open Next Edit model trained by @continue.dev

Runnable with @ollamabot.bsky.social and available today in VS Code

Continue 1.0 is here! Combining our open-source IDE extensions with hub.continue.dev makes it frictionless to use custom AI code assistants. Discover the models, rules, prompts, docs, and other building blocks you need to become an amplified developer ✨