Microsoft Unveils Phi-4-reasoning-vision-15B: A Compact Multimodal AI Model Outperforming Larger Systems with Less Compute and Data
Microsoft has released Phi-4-reasoning-vision-15B, a new open-weight multimodal AI model with 15 billion parameters. This model is designed to match or exceed the performance of much larger systems while significantly reducing compute and training data consumption. It processes both images and text, capable of complex tasks such as solving math and science problems, interpreting charts, navigating GUIs, and handling visual tasks like photo captioning. Available through Microsoft Foundry, HuggingFace, and GitHub under a permissive license, Phi-4-reasoning-vision-15B represents Microsoft's ongoing effort to demonstrate the competitiveness of carefully engineered small models against the industry's largest AI systems. A key highlight is its training on approximately 200 billion tokens of multimodal data, a fraction of what rival models typically require.
Microsoft on Tuesday released Phi-4-reasoning-vision-15B, a compact open-weight multimodal AI model. The company states that this model matches or exceeds the performance of systems many times its size, while consuming a fraction of the compute and training data. This release marks the latest and most technically ambitious chapter in Microsoft's year-long campaign to prove that carefully engineered small models can compete with, and in key areas outperform, the industry's largest AI systems.
The 15-billion-parameter model is immediately available through Microsoft Foundry, HuggingFace, and GitHub under a permissive license. It is designed to process both images and text, enabling it to reason through complex math and science problems, interpret charts and documents, navigate graphical user interfaces, and handle everyday visual tasks such as captioning photos and reading receipts. Its introduction comes at a time when the AI industry is grappling with a fundamental tension: while the biggest models deliver the best raw performance, their enormous cost, latency, and energy consumption often make them impractical for many real-world deployments.
The Microsoft Research team articulated their goal in the model's official announcement: "Our goal is to contribute practical insight to the community on building smaller, efficient multimodal reasoning models, and to share an open-weight model that is competitive with models of similar size at general vision-language tasks, excels at computer use, and excels on scientific and mathematical multimodal reasoning."
Perhaps the most striking claim in the release pertains to the relatively small amount of training data the model required compared to its competitors. Phi-4-reasoning-vision-15B was trained on approximately 200 billion tokens of multimodal data. This was built atop the Phi-4-Reasoning language backbone, which itself was trained on 16 billion tokens, and the foundational Phi-4 model, trained on 400 billion unique tokens. This contrasts sharply with rival models, which typically require significantly more data.