Microsoft has introduced a new version of its AI code model, MAI-Code-1.1-Flash, specifically designed to run directly on Windows PCs without needing internet access or cloud servers. This model, which was first introduced in GitHub Copilot earlier this summer, performs calculations locally, eliminating costs associated with token usage. The model requires 53 GB of memory storage, and Microsoft recommends more than 120 GB of graphics memory for optimal performance. It is currently available on high-end machines like the Surface Laptop Ultra, which has 128 GB of unified memory, as the entry-level version only has 24 GB. The model was released on August 11 and is developed by Microsoft AI, the division responsible for creating in-house models and setting ethical guidelines for them in September. Compared to the previous version of the model, MAI-Code-1.1-Flash is said to produce higher quality code faster and at a quarter of the cost. In GitHub Copilot, the new model costs $0.20 (about €0.18 HT) per million input tokens and $1.20 (about €1.07 HT) for output, compared to $0.75 and $4.50 (about €0.67 and €4.02 HT) for the earlier version. A token is a small fragment of a word, the unit by which AI systems read, write, and charge. Microsoft claims that running this model locally is eight to ten times cheaper than Claude Sonnet 5.5, a competing AI model. However, in terms of performance, MAI-Code-1.1-Flash scores 62.9% on Terminal-Bench 2.1, a test involving 89 terminal tasks, which is lower than scores from other models like GPT-5.6 Luna and Claude Opus 5.5. Microsoft has optimized the model to run efficiently on local machines, making it suitable for executing large AI models directly on Windows. The model contains 137 billion parameters, but only 6.8 billion are active at each step, reducing the overall size. Microsoft has also used quantization, a technique that stores parameters with less precision—about 3.3 bits on average—reducing the model’s size to 53 GB, which is 80% smaller than the cloud version. The local version of the model performs slightly better on Terminal-Bench 2.1, with 66.3% success compared to 62.9% in the cloud version, according to Microsoft's tests conducted on October 5 using a Surface Laptop Ultra. However, the model's performance drops when it uses its maximum context window of 256,000 tokens, increasing memory usage to 75.5 GB. By the end of October, GitHub Copilot will offer two experimental options in its command line tool, application, and editor VS Code. The Auto mode lets Copilot decide whether to use the local or cloud model, while developers can manually choose the local model. Both options allow commands to run in a sandbox, an isolated environment that limits access to files and the network, if enabled in project settings or using the /sandbox command. While MAI-Code-1.1-Flash can be a cost-effective option for coding tasks, it may not perform as well on complex tasks compared to other models. The local version is initially targeted at PCs with the Nvidia RTX Spark chip, such as the Surface Laptop Ultra, which is now available for pre-order and will be delivered starting October 16. The model can also run on other configurations, provided they have sufficient VRAM.