The Phi-3.5 family, capable of a staggering 128K token context length, is now available to test

Microsoft has unveiled the Phi-3.5 family of models, a trio of AI marvels designed to outshine their competitors. This lineup includes the Phi-3.5-vision, Phi-3.5-MoE, and Phi-3.5-mini, each crafted to push the boundaries of what’s possible in artificial intelligence.

These models are not just about raw power; they’re about intelligent, efficient computing, leveraging synthetic data and the vast expanse of the internet to support a whopping 128K token context length.

Moreover, they’re all available on Hugging Face under an MIT license, making them accessible to innovators everywhere.

The Phi-3.5: Everything you need to know about the new version of the groundbreaking AI family

The MoE model is a true standout, the first in its family to utilize Mixture of Experts (MoE) technology. Imagine an advanced model that activates only a fraction of its parameters to deliver top-notch performance. The Phi-3.5-MoE does that with its 16 x 3.8B MoE model, activating just 6.6B parameters. It’s a game-changer, outperforming several of its predecessors and giving the current AI leader, GPT-4o-mini, a run for its money.

Then there’s the mini version, a testament to the power of efficiency. Despite its relatively modest size, this model punches well above its weight, challenging and surpassing models with significantly more parameters. It’s a marvel in multilingual tasks, proving that sometimes, less is more. Its support for a 128K context length puts it leagues ahead of its main rival, the Gemma-2 family.

Not to be overlooked, the Phi-3.5 Vision brings a new dimension to AI with enhanced multi-frame image understanding. It’s not just about seeing; it’s about comprehending and reasoning across multiple frames, a capability to revolutionize how machines interpret visual data. With improvements across several benchmarks, the Phi-3.5 Vision sets new standards for AI’s visual capabilities.

Microsoft’s latest release is more than just a showcase of technological prowess; it signals its commitment to advancing AI in innovative and accessible ways. Just months ago, the Redmond-based tech giant released the Phi-3.0 family. Hence, the Phi-3.5 family of models represents a leap forward in AI model efficiency and capabilities, promising broader adoption across various applications.

You can get all three models on Hugging Face as follows:

• The MoE model
• The Mini model
• The Vision model