The Phi-3 models are the most capable and cost-effective Small Language Models(SLMs) available, outperforming models of the same size and the next size up across a variety of language, reasoning, coding, and math benchmarks. This release expands the selection of high-quality models for customers, offering more practical choices for composing and building generative AI applications.
The Phi-3 Family includes mini, small, medium and vision versions, trained based on different parameter amounts to serve various application scenarios. Each model is instruction-tuned and developed in accordance with Microsoft's Responsible AI, safety and security standards to ensure it's ready to use off-the-shelf. Phi-3-mini outperforms models twice its size, and Phi-3-small and Phi-3-medium outperform much larger models, including GPT-3.5T.
| Tasks | Phi-3 |
| Language Tasks | Yes |
| Math & Reasoning | Yes |
| Coding | Yes |
| Function Calling | No |
| Self Orchestration (Assistant) | No |
| Dedicated Embedding Models | No |
Phi-3-mini, a 3.8B parameter language model, is available on Microsoft Azure AI Studio, Hugging Face, and Ollama. It offers two context lengths: 128K and 4K.
Phi-3-mini is a Transformer-based language model with 3.8 billion parameters. It was trained using high-quality data containing educationally useful information, augmented with new data sources consisting of various NLP synthetic texts, and both internal and external chat datasets, which significantly improve chat capabilities. Additionally, Phi-3-mini has been chat fine-tuned after pre-training through supervised fine-tuning (SFT) and Direct Preference Optimization (DPO). Following this post-training, Phi-3-mini has demonstrated significant improvements in several capabilities, particularly in alignment, robustness, and safety. The model is part of the Phi-3 family and comes in the mini version with two variants, 4K and 128K, which represent the context length (in tokens) that it can support.
Phi-3.5 mini is a lightweight, state-of-the-art open model built upon datasets used for Phi-3 - synthetic data and filtered publicly available websites - with a focus on very high-quality, reasoning dense data. The model belongs to the Phi-3 model family and supports 128K token context length. The model underwent a rigorous enhancement process, incorporating both supervised fine-tuning, proximal policy optimization, and direct preference optimization to ensure precise instruction adherence and robust safety measures.
Phi-3.5 Mini has 3.8B parameters and is a dense decoder-only Transformer model using the same tokenizer as Phi-3 Mini.
Overall, the model with only 3.8B-param achieves a similar level of multilingual language understanding and reasoning ability as much larger models. However, it is still fundamentally limited by its size for certain tasks. The model simply does not have the capacity to store too much factual knowledge, therefore, users may experience factual incorrectness. However, we believe such weakness can be resolved by augmenting Phi-3.5 with a search engine, particularly when using the model under RAG settings.
The table below highlights multilingual capability of the Phi-3 on multilingual MMLU, MEGA, and multilingual MMLU-pro datasets. Overall, we observed that even with just 3.8B active parameters, the model is very competitive on multilingual tasks in comparison to other models with a much bigger active parameters.
Phi-3-small, a 7B parameter language model, available in two context lengths 128K and 8K. outperforms GPT-3.5T across a variety of language, reasoning, coding, and math benchmarks.
Phi-3-small is a Transformer-based language model with 7 billion parameters. It was trained using high-quality data containing educationally useful information, augmented with new data sources that consist of various NLP synthetic texts, and both internal and external chat datasets, which significantly improve chat capabilities. In addition, Phi-3-small has been chat fine-tuned after pre-training via supervised fine-tuning (SFT) and Direct Preference Optimization (DPO). Following this post-training, Phi-3-small has shown significant improvements in several capabilities, particularly in alignment, robustness, and safety. Phi-3-small is also more intensively trained on multilingual datasets compared to Phi-3-Mini. The model family offers two variants, 8K and 128K, which represent the context length (in tokens) that it can support.
Phi-3-medium, a 14B parameter language model, available in two context lengths 128K and 4K., continues the trend by outperforming Gemini 1.0 Pro.
Phi-3-medium is a Transformer-based language model with 14 billion parameters. It was trained using high-quality data containing educationally useful information, augmented with new data sources that consist of various NLP synthetic texts, and both internal and external chat datasets, which significantly improve chat capabilities. Additionally, Phi-3-medium has been chat fine-tuned after pre-training through supervised fine-tuning (SFT) and Direct Preference Optimization (DPO). Following this post-training, Phi-3-medium has exhibited significant improvements in several capabilities, particularly in alignment, robustness, and safety. The model family offers two variants, 4K and 128K, which represent the context length (in tokens) that it can support.
[!NOTE] We recommend switching to Phi-3.5-MoE as an upgrade of Phi-3-medium as the MoE model is a much better and cost effective model.
The Phi-3-vision, a 4.2B parameter multimodal model with language and vision capabilities, outperforms larger models like Claude-3 Haiku and Gemini 1.0 Pro V in general visual reasoning, OCR, and table and chart understanding tasks.
Phi-3-vision is the first multimodal model in the Phi-3 family, bringing together text and images. Phi-3-vision can be used to reason over real-world images and extract and reason over text from images. It has also been optimized for chart and diagram understanding and can be used to generate insights and answer questions. Phi-3-vision builds on the language capabilities of the Phi-3-mini, continuing to pack strong language and image reasoning quality in a small size.
Phi-3.5 Vision is a lightweight, state-of-the-art open multimodal model built upon datasets which include - synthetic data and filtered publicly available websites - with a focus on very high-quality, reasoning dense data both on text and vision. The model belongs to the Phi-3 model family, and the multimodal version comes with 128K context length (in tokens) it can support. The model underwent a rigorous enhancement process, incorporating both supervised fine-tuning and direct preference optimization to ensure precise instruction adherence and robust safety measures.
Phi-3.5 Vision has 4.2B parameters and contains image encoder, connector, projector, and Phi-3 Mini language model.
The model is intended for broad commercial and research use in English. The model provides uses for general purpose AI systems and applications with visual and text input capabilities which require
- memory/compute constrained environments.
- latency bound scenarios.
- general image understanding.
- OCR
- chart and table understanding.
- multiple image comparison. 7)multi-image or video clip summarization.
Phi-3.5-vision model is designed to accelerate research on efficient language and multimodal models, for use as a building block for generative AI powered features
Phi-3.5 MoE is a lightweight, state-of-the-art open model built upon datasets used for Phi-3 - synthetic data and filtered publicly available documents - with a focus on very high-quality, reasoning dense data. The model supports multilingual and comes with 128K context length (in tokens). The model underwent a rigorous enhancement process, incorporating supervised fine-tuning, proximal policy optimization, and direct preference optimization to ensure precise instruction adherence and robust safety measures.
Phi-3 MoE has 16x3.8B parameters with 6.6B active parameters when using 2 experts. The model is a mixture-of-expert decoder-only Transformer model using the tokenizer with vocabulary size of 32,064.
The model is intended for broad commercial and research use in English. The model provides uses for general purpose AI systems and applications which require.
- memory/compute constrained environments.
- latency bound scenarios.
- strong reasoning (especially math and logic).
The MoE model is designed to accelerate research on language and multimodal models, for use as a building block for generative AI powered features and requires additional compute resources.
Note
Phi-3 models do not perform as well on factual knowledge benchmarks (such as TriviaQA) as the smaller model size results in less capacity to retain facts.
We are introducing Phi Silica which is built from the Phi series of models and is designed specifically for the NPUs in Copilot+ PCs. Windows is the first platform to have a state-of-the-art small language model (SLM) custom built for the NPU and shipping inbox. Phi Silica API along with OCR, Studio Effects, Live Captions, and Recall User Activity APIs will be available in Windows Copilot Library in June. More APIs like Vector Embedding, RAG API, and Text Summarization will be coming later.
The primary difference between the two ONNX models, “cpu-int4-rtn-block-32” and “cpu-int4-rtn-block-32-acc-level-4”, is the accuracy level. The model with “acc-level-4” is designed to balance latency versus accuracy, with a minor trade-off in accuracy for better performance, which might be particularly suitable for mobile devices
| Customer Need | Task | Start with | More Details |
| Need a model that simply summarizes a thread of messages | Conversation Summarization | Phi-3 text model | Deciding factor here is that the customer has a well defined and straight forward language task |
| A free math tutor app for kids | Math and Reasoning | Phi-3 text models | Because the app is free customers want a solution that does not cost them on a recurring basis |
| Self Patrol Car Camera | Vision analysis | Phi-Vision | Need a solution that can work on edge without internet |
| Wants to build an AI based travel booking agent | Needs complex planning, function calling and orchestration | GPT models | Need ability to plan, call APIs to gather information and execute |
| Wants to build a copilot for their employees | RAG, multiple domain, complex and open ended | GPT models | Open ended scenario, needs broader world knowledge, hence a larger model is more suited |