In this article

We are excited to present Phi-3, a household of unfastened AI models developed by Microsoft. Phi-3 models are the most susceptible and cost-effective tiny connection models (SLMs) available, outperforming models of the aforesaid size and adjacent size up crossed a assortment of language, reasoning, coding, and mathematics benchmarks. This merchandise expands the enactment of high-quality models for customers, offering much applicable choices arsenic they constitute and physique generative AI applications.

Starting today, Phi-3-mini, a 3.8B connection exemplary is disposable connected Microsoft Azure AI Studio, Hugging Face, and Ollama. 

• Phi-3-mini is disposable successful 2 context-length variants—4K and 128K tokens. It is the archetypal exemplary successful its people to enactment a discourse model of up to 128K tokens, with small interaction connected quality.
• It is instruction-tuned, meaning that it’s trained to travel antithetic types of instructions reflecting however radical usually communicate. This ensures the exemplary is acceptable to usage out-of-the-box.
• It is disposable connected Azure AI to instrumentality vantage of the deploy-eval-finetune toolchain, and is disposable connected Ollama for developers to tally locally connected their laptops.
• It has been optimized for ONNX Runtime with enactment for Windows DirectML on with cross-platform enactment crossed graphics processing portion (GPU), CPU, and adjacent mobile hardware.
• It is besides disposable arsenic an NVIDIA NIM microservice with a modular API interface that tin beryllium deployed anywhere. And has been optimized for NVIDIA GPUs. 

In the coming weeks, further models volition beryllium added to Phi-3 household to connection customers adjacent much flexibility crossed the quality-cost curve. Phi-3-small (7B) and Phi-3-medium (14B) volition beryllium disposable successful the Azure AI exemplary catalog and different exemplary gardens shortly.   

Microsoft continues to connection the champion models crossed the quality-cost curve and today’s Phi-3 merchandise expands the enactment of models with state-of-the-art tiny models.

Groundbreaking show astatine a tiny size

Phi-3 models importantly outperform connection models of the aforesaid and larger sizes connected cardinal benchmarks (see benchmark numbers below, higher is better). Phi-3-mini does amended than models doubly its size, and Phi-3-small and Phi-3-medium outperform overmuch larger models, including GPT-3.5T.  

All reported numbers are produced with the aforesaid pipeline to guarantee that the numbers are comparable. As a result, these numbers whitethorn disagree from different published numbers owed to flimsy differences successful the valuation methodology. More details connected benchmarks are provided successful our technical paper. 

Note: Phi-3 models bash not execute arsenic good connected factual cognition benchmarks (such arsenic TriviaQA) arsenic the smaller exemplary size results successful little capableness to clasp facts. 

Safety-first exemplary design

Phi-3 models were developed successful accordance with the Microsoft Responsible AI Standard, which is simply a company-wide acceptable of requirements based connected the pursuing six principles: accountability, transparency, fairness, reliability and safety, privateness and security, and inclusiveness. Phi-3 models underwent rigorous information measurement and evaluation, red-teaming, delicate usage review, and adherence to information guidance to assistance guarantee that these models are responsibly developed, tested, and deployed successful alignment with Microsoft’s standards and champion practices.  

Building connected our anterior enactment with Phi models (“Textbooks Are All You Need”), Phi-3 models are besides trained utilizing high-quality data. They were further improved with extended information post-training, including reinforcement learning from quality feedback (RLHF), automated investigating and evaluations crossed dozens of harm categories, and manual red-teaming. Our attack to information grooming and evaluations are elaborate successful our technical paper, and we outline recommended uses and limitations successful the exemplary cards. See the model paper collection. 

Unlocking caller capabilities

Microsoft’s acquisition shipping copilots and enabling customers to alteration their businesses with generative AI utilizing Azure AI has highlighted the increasing request for different-size models crossed the quality-cost curve for antithetic tasks. Small connection models, similar Phi-3, are particularly large for: 

• Resource constrained environments including on-device and offline inference scenarios.
• Latency bound scenarios wherever accelerated effect times are critical.
• Cost constrained usage cases, peculiarly those with simpler tasks.

For much connected tiny connection models, spot our Microsoft Source Blog.

Thanks to their smaller size, Phi-3 models tin beryllium utilized successful compute-limited inference environments. Phi-3-mini, successful particular, tin beryllium utilized on-device, particularly erstwhile further optimized with ONNX Runtime for cross-platform availability. The smaller size of Phi-3 models besides makes fine-tuning oregon customization easier and much affordable. In addition, their little computational needs marque them a little outgo enactment with overmuch amended latency. The longer discourse model enables taking successful and reasoning implicit ample substance content—documents, web pages, code, and more. Phi-3-mini demonstrates beardown reasoning and logic capabilities, making it a bully campaigner for analytical tasks. 

Customers are already gathering solutions with Phi-3. One illustration wherever Phi-3 is already demonstrating worth is successful agriculture, wherever net mightiness not beryllium readily accessible. Powerful tiny models similar Phi-3 on with Microsoft copilot templates are disposable to farmers astatine the constituent of request and supply the further payment of moving astatine reduced cost, making AI technologies adjacent much accessible.  

ITC, a starring concern conglomerate based successful India, is leveraging Phi-3 arsenic portion of their continued collaboration with Microsoft connected the copilot for Krishi Mitra, a farmer-facing app that reaches implicit a cardinal farmers.

“Our extremity with the Krishi Mitra copilot is to amended ratio portion maintaining the accuracy of a ample connection model. We are excited to spouse with Microsoft connected utilizing fine-tuned versions of Phi-3 to conscionable some our goals—efficiency and accuracy!”   

Saif Naik, Head of Technology, ITCMAARS

Originating successful Microsoft Research, Phi models person been broadly used, with Phi-2 downloaded implicit 2 cardinal times. The Phi bid of models person achieved singular show with strategical information curation and innovative scaling. Starting with Phi-1, a exemplary utilized for Python coding, to Phi-1.5, enhancing reasoning and understanding, and past to Phi-2, a 2.7 billion-parameter exemplary outperforming those up to 25 times its size successful connection comprehension.¹ Each iteration has leveraged high-quality grooming information and cognition transportation techniques to challenge conventional scaling laws. 

Get started today

To acquisition Phi-3 for yourself, commencement with playing with the exemplary connected Azure AI Playground. You tin besides find the exemplary connected the Hugging Chat playground. Start gathering with and customizing Phi-3 for your scenarios utilizing the Azure AI Studio. Join america to larn much astir Phi-3 during a special live stream of the AI Show.  

---

¹Microsoft Research Blog, Phi-2: The astonishing powerfulness of tiny connection models, December 12, 2023.

The station Introducing Phi-3: Redefining what’s imaginable with SLMs appeared archetypal connected Microsoft Azure Blog.