Microsoft Introduces Multimodal SLMs Trained on NVIDIA GPUs

Microsoft has introduced the newest additions to its Phi household of small language fashions (SLMs), that includes the brand new Phi-4-multimodal and Phi-4-mini fashions, each skilled utilizing NVIDIA GPUs. This improvement marks a big step within the evolution of language fashions, specializing in effectivity and flexibility, in accordance with NVIDIA.

Developments in Small Language Fashions

SLMs have emerged as a sensible resolution to the challenges posed by giant language fashions (LLMs), which, regardless of their capabilities, require substantial computational sources. SLMs are designed to function effectively inside constrained environments, making them appropriate for deployment on gadgets with restricted reminiscence and computational energy.

Microsoft’s new Phi-4-multimodal mannequin is especially noteworthy for its capacity to course of a number of sorts of knowledge, together with textual content, audio, and pictures. This functionality opens up new prospects for purposes resembling automated speech recognition, translation, and visible reasoning. The mannequin’s coaching concerned 512 NVIDIA A100-80GB GPUs over 21 days, underscoring the intensive computational efforts required to realize its capabilities.

Phi-4-multimodal and Phi-4-mini

The Phi-4-multimodal mannequin boasts 5.6 billion parameters and has demonstrated superior efficiency in automated speech recognition, rating first on the Huggingface OpenASR leaderboard with a phrase error price of 6.14%. This achievement highlights the mannequin’s potential in enhancing speech recognition applied sciences.

Alongside Phi-4-multimodal, Microsoft additionally launched Phi-4-mini, a text-only mannequin optimized for chat purposes. With 3.8 billion parameters, Phi-4-mini is designed to deal with long-form content material effectively, providing a context window of 128K tokens. Its coaching concerned 1024 NVIDIA A100 80GB GPUs over 14 days, reflecting the mannequin’s deal with high-quality academic knowledge and code.

Deployment and Accessibility

Each fashions can be found on Microsoft’s Azure AI Foundry, offering a platform for designing, customizing, and managing AI purposes. Customers may also discover these fashions by way of the NVIDIA API Catalog, which affords a sandbox setting for testing and integrating these fashions into varied purposes.

NVIDIA’s collaboration with Microsoft extends past simply coaching these fashions. The partnership consists of optimizing software program and fashions like Phi to advertise AI transparency and help open-source initiatives. This collaboration goals to advance AI expertise throughout industries, from healthcare to life sciences.

For extra detailed data, go to the NVIDIA weblog.

Picture supply: Shutterstock