25/08/2026
Avec Junie Local, JetBrains propose désormais une approche locale de l’IA : le modèle s’exécute directement sur votre Mac, sans envoi de vos prompts, de votre code source ou de vos diffs vers le réseau.
🔒 Plus de confidentialité
⚡ Des performances optimisées pour le développement
💰 Aucun coût à l’usage ni quota de tokens
🛠️ Une configuration pensée pour fonctionner simplement
Une approche particulièrement intéressante pour les équipes qui travaillent avec des données sensibles ou sous NDA.
👉 Découvrez comment fonctionne Junie Local et les choix techniques qui se cachent derrière cette nouvelle expérience.
Junie can now run entirely on your machine, and it is free.
Junie has connected to local runtimes like Ollama and LM Studio for a while. The difference with Junie Local is that we chose the model and then optimized the entire stack around it: the agent harness, the model parameters, and the inference engine.
Run /local inside Junie. The model downloads, a local server starts, and the agent switches over. Nothing to configure. After the download, your prompts, source, and diffs never touch the network.
On JetBrains' private test set, the model scores on par with Sonnet 4.5 and a few points behind GPT-5. Ours runs with reasoning disabled, because it added little quality for two to three times the tokens spent.
A couple of the decisions behind it: We start with M5 – its Neural Accelerator has 8-bit arithmetic instructions that gave us roughly 40% more prefill throughput once we patched MLX-VLM to use them, and prefill is where a coding agent actually spends its time. We opted for Qwen3.6 rather than 3.8 because Qwen 3.8 requires reasoning to work reliably, which makes tasks around four times slower.
For privacy-conscious teams and those bound by client NDAs, this removes the vendor from the conversation entirely. For everyone else, nothing is metered, so the long mechanical work you never delegated becomes worth delegating.
What Junie Local needs: an M5 Mac with 64 GB of RAM and 40 GB of free disk space.
We know that it's a lot to ask, but it is just the starting point, not the finish line. We have prototypes running on DGX Spark and RTX 5090, and we are looking at 24 GB cards and more models.
Learn more: https://blog.jetbrains.com/junie/2026/08/junie-local-launch/
