Introducing Phi-3: Redefining what’s potential with SLMs


We’re excited to introduce Phi-3, a household of open AI fashions developed by Microsoft. Phi-3 fashions are the most succesful and cost-effective small language fashions (SLMs) out there, outperforming fashions of the identical measurement and subsequent measurement up throughout quite a lot of language, reasoning, coding, and math benchmarks. This launch expands the number of high-quality fashions for patrons, providing extra sensible selections as they compose and construct generative AI functions.

Beginning at present, Phi-3-mini, a 3.8B language mannequin is out there on Microsoft Azure AI Studio, Hugging Face, and Ollama

  • Phi-3-mini is out there in two context-length variants—4K and 128K tokens. It’s the first mannequin in its class to help a context window of as much as 128K tokens, with little affect on high quality.
  • It’s instruction-tuned, that means that it’s skilled to comply with several types of directions reflecting how individuals usually talk. This ensures the mannequin is able to use out-of-the-box.
  • It’s out there on Azure AI to reap the benefits of the deploy-eval-finetune toolchain, and is out there on Ollama for builders to run regionally on their laptops.
  • It has been optimized for ONNX Runtime with help for Windows DirectML together with cross-platform help throughout graphics processing unit (GPU), CPU, and even cell {hardware}.
  • It’s also out there as an NVIDIA NIM microservice with a normal API interface that may be deployed wherever. And has been optimized for NVIDIA GPUs

Within the coming weeks, further fashions will probably be added to Phi-3 household to supply prospects much more flexibility throughout the quality-cost curve. Phi-3-small (7B) and Phi-3-medium (14B) will probably be out there within the Azure AI mannequin catalog and different mannequin gardens shortly.   

Microsoft continues to supply one of the best fashions throughout the quality-cost curve and at present’s Phi-3 launch expands the number of fashions with state-of-the-art small fashions.

abstract image

Azure AI Studio

Phi-3-mini is now out there

Groundbreaking efficiency at a small measurement

Phi-3 fashions considerably outperform language fashions of the identical and bigger sizes on key benchmarks (see benchmark numbers under, increased is healthier). Phi-3-mini does higher than fashions twice its measurement, and Phi-3-small and Phi-3-medium outperform a lot bigger fashions, together with GPT-3.5T.  

All reported numbers are produced with the identical pipeline to make sure that the numbers are comparable. Consequently, these numbers might differ from different revealed numbers on account of slight variations within the analysis methodology. Extra particulars on benchmarks are supplied in our technical paper

Notice: Phi-3 fashions don’t carry out as nicely on factual information benchmarks (equivalent to TriviaQA) because the smaller mannequin measurement ends in much less capability to retain info. 

Security-first mannequin design

Phi-3 fashions have been developed in accordance with the Microsoft Responsible AI Standard, which is a company-wide set of necessities primarily based on the next six rules: accountability, transparency, equity, reliability and security, privateness and safety, and inclusiveness. Phi-3 fashions underwent rigorous security measurement and analysis, red-teaming, delicate use evaluation, and adherence to safety steering to assist make sure that these fashions are responsibly developed, examined, and deployed in alignment with Microsoft’s requirements and finest practices.  

Constructing on our prior work with Phi fashions (“Textbooks Are All You Need”), Phi-3 fashions are additionally skilled utilizing high-quality information. They have been additional improved with intensive security post-training, together with reinforcement studying from human suggestions (RLHF), automated testing and evaluations throughout dozens of hurt classes, and handbook red-teaming. Our method to security coaching and evaluations are detailed in our technical paper, and we define really useful makes use of and limitations within the mannequin playing cards. See the model card collection. 

Unlocking new capabilities

Microsoft’s expertise transport copilots and enabling prospects to rework their companies with generative AI utilizing Azure AI has highlighted the rising want for different-size fashions throughout the quality-cost curve for various duties. Small language fashions, like Phi-3, are particularly nice for: 

  • Useful resource constrained environments together with on-device and offline inference situations.
  • Latency certain situations the place quick response instances are crucial.
  • Value constrained use instances, significantly these with less complicated duties.

For extra on small language fashions, see our Microsoft Source Blog.

Because of their smaller measurement, Phi-3 fashions can be utilized in compute-limited inference environments. Phi-3-mini, particularly, can be utilized on-device, particularly when additional optimized with ONNX Runtime for cross-platform availability. The smaller measurement of Phi-3 fashions additionally makes fine-tuning or customization simpler and extra reasonably priced. As well as, their decrease computational wants make them a decrease price possibility with significantly better latency. The longer context window allows taking in and reasoning over massive textual content content material—paperwork, internet pages, code, and extra. Phi-3-mini demonstrates robust reasoning and logic capabilities, making it a great candidate for analytical duties. 

Prospects are already constructing options with Phi-3. One instance the place Phi-3 is already demonstrating worth is in agriculture, the place web won’t be readily accessible. Highly effective small fashions like Phi-3 together with Microsoft copilot templates can be found to farmers on the level of want and supply the extra good thing about working at lowered price, making AI applied sciences much more accessible.  

ITC, a number one enterprise conglomerate primarily based in India, is leveraging Phi-3 as a part of their continued collaboration with Microsoft on the copilot for Krishi Mitra, a farmer-facing app that reaches over 1,000,000 farmers.

Our purpose with the Krishi Mitra copilot is to enhance effectivity whereas sustaining the accuracy of a giant language mannequin. We’re excited to accomplice with Microsoft on utilizing fine-tuned variations of Phi-3 to fulfill each our targets—effectivity and accuracy!”   

Saif Naik, Head of Expertise, ITCMAARS

Originating in Microsoft Analysis, Phi fashions have been broadly used, with Phi-2 downloaded over 2 million instances. The Phi collection of fashions have achieved exceptional efficiency with strategic information curation and progressive scaling. Beginning with Phi-1, a mannequin used for Python coding, to Phi-1.5, enhancing reasoning and understanding, after which to Phi-2, a 2.7 billion-parameter mannequin outperforming these as much as 25 instances its measurement in language comprehension.1 Every iteration has leveraged high-quality coaching information and information switch methods to problem standard scaling legal guidelines. 

Get began at present

To expertise Phi-3 for your self, begin with taking part in with the mannequin on Azure AI Playground. You too can discover the mannequin on the Hugging Chat playground. Begin constructing with and customizing Phi-3 to your situations utilizing the Azure AI Studio. Be part of us to be taught extra about Phi-3 throughout a particular live stream of the AI Present.  


1 Microsoft Research Blog, Phi-2: The surprising power of small language models, December 12, 2023.



Leave a Reply

Your email address will not be published. Required fields are marked *