MAY-Code-1-Flash

We introduce: MAI-Code-1-Flash

Today we introduce MAI code 1 flash, a new Microsoft-Codation model developed for fast and efficient support in everyday developer workflows. It is created by Microsoft consistently using clean and appropriately licensed data. The model will be GitHub Copilot single user in Visual Studio Code provided in the model selection and under the standard automatic selection. Functions and skills

- Agent encoding in real developer environments, trained and designed for GitHub copilot use to work together better. - Adaptive thinking, stays concise with simple enquiries and spends more thinking budget for complex tasks. - Strong follow instructions in single-turn and multi-turn scenarios. MAI Code-1 flash has been developed with the simple aim of providing high-quality coding aid with higher efficiency. It surpasses Claude Haiku 4.5 with a better price-performance ratio for all coding benchmarks. Designed for developers, not for benchmarks

Coding models are most useful if they provide good performance in the same environment that developers use daily. For this reason, we have developed MAI code 1 flash with production processes and not only optimized for benchmarks. The model was trained directly with the GitHub copilot cable harnesses used in production. As a result, it can learn how it interacts with surrounding tools and systems in agent coding tasks, making it particularly suitable for real copilot workflows compared to other available models. During the training, we evaluated test points for core tasks of software development, answering repository questions, refactoring and telemetry-based tasks adapted to the real use of GitHub copilot. This coordination between training, evaluation and production contributes to real developer quality offline improvements. Developed to maximize value per token

MAI-Code-1-Flash was trained with adaptive solution length control, which helps the model to adjust the depth of its reaction to the task. For simpler requests, it may remain concise and issue more argumentation budget if a problem requires a deeper analysis or more comprehensive code changes. In practice, this means that developers see useful results earlier. We see that MAI code-1 flash solves more difficult problems with up to 60% less tokens. This helps to reduce latency, reduce costs, improve return on tokens and make interactive workflows feel more smooth. Benchmark results in production

To understand both quality and efficiency, we compared MAI code-1 flash with Claude Haiku 4.5 on SWE-Bench Verified, SWE-Bench Pro, SWE-Bench Multilingual and Terminal Bench 2, using the same production system that developers use for their everyday coding tasks. We have measured the success and the average number of solution tokens required to complete each task. MAI-Code-1-Flash surpasses Claude Haiku 4.5 in all tested core-coding benchmarks with higher success rates for all four ratings, including a lead of +16 points in the various real tasks of SWE-Bench Pro (51.2 % versus 35.2 %). It is not only more intelligent; It is slimmer and solves more difficult problems with up to 60% less tokens at SWE-Bench Verified, which proves that higher accuracy and higher efficiency are no longer a compromise. Tasks in mathematics, natural sciences, follow instructions and agent encoding

MAI-Code-1 flash has the nose in front of all benchmarks in the table, with the largest lead in the precise IF-Bench-Instruction sequence (+28.9) and the closest in the Advanced IF (+14.5) based on headings. The strict compliance of instructions is transferred to the use of agent tools. In addition, MAI-Code-1-Flash Claude Haiku-4.5 also surpasses the core competencies of thinking in mathematics, natural sciences and visual generation. Standard benchmarks reward memorization as well as logical thinking. For example, a model that has seen the Monty Hall problem will answer it correctly, but the prices reverse and it fails. We have created a benchmark with 186 questions and 34 categories around opposing traps such as reversed classics, impossible tasks and under-determined scenarios to see whether the models were actually consequent or only model comparisons. MAI-Code-1-Flash surpasses Claude Haiku as a whole with 4.5 and achieves an adjusted accuracy of 85.8%, with particularly strong performance in thinking, following instructions and recognizing impossible problems. We also see room for the growth of the model, as the accuracy of central Gegner categories such as setting traps remained below 50%. Try it out

MAI-Code-1-Flash is now available for individual users of VS Code GitHub Copilot. No additional equipment is required. As the rollout progresses, you can see how GitHub copilot will forward tasks via the automatic selection of MAI code 1 flash or that the model is available directly in the model selection. Here are some entertaining example apps that we created with MAI Code 1 flash in VS Code:

Build the future with us

We are a slim, fast-paced laboratory that consists of some of the most talented minds in the world. We have an exciting computing roadmap at MAI: Our next-generation GB200 cluster is now operational. And we have an ambitious mission we really believe in. We are also lucky to work with incredible product teams that give our models the opportunity to reach billions of users and achieve an enormous positive effect. If you are a brilliant, very ambitious person with little ego, fit us exactly – come to us while we work on our next generation of models!