#What New Capabilities Do Microsoft's AI Models Offer?
Microsoft Research has recently launched a groundbreaking suite of AI models that surpass current offerings from major industry players such as OpenAI and Google. Known as the Fara1.5 family, the models were announced on May 22, featuring three versions with parameter counts of 4 billion, 9 billion, and 27 billion.
The flagship model, which has 27 billion parameters, achieved an impressive score of 72% on the Online-Mind2Web benchmark. This benchmark tests AI agents against a range of web tasks across 136 live websites. In comparison, OpenAI's Operator scored 58.3%, while Google’s Gemini 2.5 Computer Use only managed 57.3%. This highlights a significant gap in performance, with Microsoft's model successfully completing nearly 75% of applicable real-world tasks.
The 9B model also showed notably strong performance with a score of 63.4%, placing it ahead of both OpenAI's and Google's models. For reference, Microsoft's previous AI model, the Fara-7B, only achieved a score of 34.1% on the same benchmark about six months prior to this release, signifying rapid advancements in performance.
#What Underlies This Enhanced Performance?
The Fara1.5 models utilize the innovative Qwen3.5 architecture and employ a feature called MagenticLite, which is a sandboxed browser interface. This environment allows AI agents to interact with web pages securely. Additionally, these models incorporate an observe-think-act loop, which includes a safeguard mechanism to enhance reliability. This system allows the AI to pause and seek user confirmation before executing critical actions, such as making purchases or altering account settings.
Microsoft has made the 9B model available on Microsoft Foundry, with plans for the other versions to follow shortly.
#What Does Open-Weight Mean for Developers?
The open-weight nature of the Fara1.5 models is significant. Unlike proprietary systems such as OpenAI's Operator and Google’s Gemini 2.5, developers can download, modify, and deploy these models freely on their hardware. Microsoft has designed the Fara1.5 family to run efficiently, even on moderate hardware, while scaling performance benefits as the model size increases. Moreover, the training pipeline has taken a notable leap forward with the introduction of FaraGen1.5, designed to provide high-quality synthetic data that improves training efficiency.
#How Might These Models Impact Crypto and DeFi Usage?
Although Microsoft did not specifically design the Fara1.5 models for cryptocurrency, their strengths could significantly influence decentralized finance (DeFi) applications. Since DeFi platforms operate as web applications, the capabilities of the Fara1.5 models in handling complex web interactions could streamline tasks such as token swaps and managing assets across various chains—activities that require precise, multi-step workflows.
The human-in-the-loop design offers an important safety feature for DeFi transactions, which are typically irreversible. The ability of an agent to pause and ask for user confirmation before executing a transaction mitigates risks associated with approving potentially harmful contracts or transferring funds incorrectly. This makes the Fara1.5 models not just a leap forward in AI's web capabilities but also a potential boon for users navigating the complexities of decentralized finance.