Microsoft has released the latest batch of open-source small AI models in the Phi series called Phi-3.5. The company has claimed that the three small language models beat competitors including Google’s Gemini 1.5 Flash, Meta’s Llama 3.1, and even OpenAI’s GPT-4o in some benchmarks.
The set of three new Phi-3.5 models include the 3.82 billion parameter Phi-3.5-mini-instruct, the 41.9 billion parameter Phi-3.5-MoE-instruct, and the 4.15 billion parameter Phi-3.5-vision-instruct, which were all designed for tasks like basic and fast reasoning, more powerful reasoning and vision tasks like image and video analysis, respectively.
All of these three models are available for download for free and can be run using a local tool like Ollama.
Despite its tiny size, the Phi-3.5 Mini Instruct model can process images as well as text and is multilingual too.
(For top technology news of the day, subscribe to our tech newsletter Today’s Cache)
The model performed quite well at the reasoning tasks and was only beaten by GPT-4o-mini out of its rivals. It also performed well on math benchmarks, significantly passing other models like Llama and Gemini.
The Phi-3.5-MoE model is the first model in this class and can combine multiple different model types into one, so each specialises in different tasks. The model was able to beat Gemini Flash 1.5, which is used in the free version of the Gemini chatbot.
It also has a comfortably large 128k context window which although significantly smaller than Gemini is equal to ChatGPT and Claude.
While the smallest model in the range was trained on 3.4 trillion tokens of data using 512 Nvidia H100 GPUs over 10 days, the MoE model used 4.9 trillion tokens and was trained over 23 days.
The major advantage of using small language models in this size is that they can be used in an app or even installed on IoT device easily because of how little compute it draws.
