Read much announcements from Azure astatine Microsoft Build 2024: New ways Azure helps you physique transformational AI experiences and The caller epoch of compute powering Azure AI solutions.

---

At Microsoft Build 2024, we are excited to adhd caller models to the Phi-3 household of small, unfastened models developed by Microsoft. We are introducing Phi-3-vision, a multimodal exemplary that brings unneurotic connection and imaginativeness capabilities. You tin try Phi-3-vision today.

Phi-3-small and Phi-3-medium, announced earlier, are present disposable connected Microsoft Azure, empowering developers with models for generative AI applications that necessitate beardown reasoning, constricted compute, and latency bound scenarios. Lastly, previously disposable Phi-3-mini, arsenic good arsenic Phi-3-medium, are present besides disposable done Azure AI’s models arsenic a work offering, allowing users to get started rapidly and easily.

The Phi-3 family

Phi-3 models are the astir 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. They are trained utilizing precocious prime grooming data, arsenic explained successful Tiny but mighty: The Phi-3 tiny connection models with large potential. The availability of Phi-3 models expands the enactment of high-quality models for Azure customers, offering much applicable choices arsenic they constitute and physique generative AI applications.

There are 4 models successful the Phi-3 exemplary family; each exemplary is instruction-tuned and developed successful accordance with Microsoft’s liable AI, safety, and information standards to guarantee it’s acceptable to usage off-the-shelf.

• Phi-3-vision is simply a 4.2B parameter multimodal exemplary with connection and imaginativeness capabilities.
• Phi-3-mini is simply a 3.8B parameter connection model, disposable successful 2 discourse lengths (128K and 4K).
• Phi-3-small is simply a 7B parameter connection model, disposable successful 2 discourse lengths (128K and 8K).
• Phi-3-medium is simply a 14B parameter connection model, disposable successful 2 discourse lengths (128K and 4K).

Find each Phi-3 models connected Azure AI and Hugging Face.

Phi-3 models person been optimized to tally crossed a assortment of hardware. Optimized variants are disposable with ONNX Runtime and DirectML providing developers with enactment crossed a wide scope of devices and platforms including mobile and web deployments. Phi-3 models are besides disposable arsenic NVIDIA NIM inference microservices with a modular API interface that tin beryllium deployed anyplace and person been optimized for inference connected NVIDIA GPUs and Intel accelerators.

It’s inspiring to spot however developers are utilizing Phi-3 to bash unthinkable things—from ITC, an Indian conglomerate, which has built a copilot for Indian farmers to inquire questions astir their crops successful their ain vernacular, to the Khan Academy, who is presently leveraging Azure OpenAI Service to powerfulness their Khanmigo for teachers aviator and experimenting with Phi-3 to amended mathematics tutoring successful an affordable, scalable, and adaptable manner. Healthcare bundle institution Epic is looking to besides usage Phi-3 to summarize analyzable diligent histories much efficiently. Seth Hain, elder vice president of R&D astatine Epic explains, “AI is embedded straight into Epic workflows to assistance lick important issues similar clinician burnout, staffing shortages, and organizational fiscal challenges. Small connection models, similar Phi-3, person robust yet businesslike reasoning capabilities that alteration america to connection high-quality generative AI astatine a little outgo crossed our applications that assistance with challenges similar summarizing analyzable diligent histories and responding faster to patients.”

Digital Green, utilized by much than 6 cardinal farmers, is introducing video to their AI assistant, Farmer.Chat, adding to their multimodal conversational interface. “We’re excited astir leveraging Phi-3 to summation the ratio of Farmer.Chat and to alteration agrarian communities to leverage the powerfulness of AI to uplift themselves,” said Rikin Gandhi, CEO, Digital Green.

Bringing multimodality to Phi-3

Phi-3-vision is the archetypal multimodal exemplary successful the Phi-3 family, bringing unneurotic substance and images, and the quality to crushed implicit real-world images and extract and crushed implicit substance from images. It has besides been optimized for illustration and diagram knowing and tin beryllium utilized to make insights and reply questions. Phi-3-vision builds connected the connection capabilities of the Phi-3-mini, continuing to battalion beardown connection and representation reasoning prime successful a tiny model.

Phi-3-vision tin make insights from charts and diagrams:

Groundbreaking show astatine a tiny size

As previously shared, Phi-3-small and Phi-3-medium outperform connection models of the aforesaid size arsenic good arsenic those that are overmuch larger.

• Phi-3-small with lone 7B parameters beats GPT-3.5T crossed a assortment of language, reasoning, coding, and mathematics benchmarks.¹
• The Phi-3-medium with 14B parameters continues the inclination and outperforms Gemini 1.0 Pro.²
• Phi-3-vision with conscionable 4.2B parameters continues that inclination and outperforms larger models specified arsenic Claude-3 Haiku and Gemini 1.0 Pro V crossed wide ocular reasoning tasks, OCR, table, and illustration knowing tasks.³

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.

See elaborate benchmarks successful the footnotes of this post.

Prioritizing safety

Phi-3 models were developed successful accordance with the Microsoft Responsible AI Standard and 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.

Phi-3 models are besides trained utilizing high-quality information and were further improved with 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 model cards.

Finally, developers utilizing the Phi-3 exemplary household tin besides instrumentality vantage of a suite of tools disposable successful Azure AI to assistance them physique safer and much trustworthy applications.

Choosing the close model

With the evolving scenery of disposable models, customers are progressively looking to leverage aggregate models successful their applications depending connected usage lawsuit and concern needs. Choosing the close exemplary depends connected the needs of a circumstantial usage case.

Small connection models are designed to execute good for simpler tasks, are much accessible and easier to usage for organizations with constricted resources, and they tin beryllium much easy fine-tuned to conscionable circumstantial needs. They are good suited for applications that request to tally locally connected a device, wherever a task doesn’t necessitate extended reasoning and a speedy effect is needed.

The prime betwixt utilizing Phi-3-mini, Phi-3-small, and Phi-3-medium depends connected the complexity of the task and disposable computational resources. They tin beryllium employed crossed a assortment of connection knowing and procreation tasks specified arsenic contented authoring, summarization, question-answering, and sentiment analysis. Beyond accepted connection tasks these models person beardown reasoning and logic capabilities, making them bully candidates for analytical tasks. The longer discourse model disposable crossed each models enables taking successful and reasoning implicit ample substance content—documents, web pages, code, and more.

Phi-3-vision is large for tasks that necessitate reasoning implicit representation and substance together. It is particularly bully for OCR tasks including reasoning and Q&A implicit extracted text, arsenic good arsenic chart, diagram, and array knowing tasks.

Get started today

To acquisition Phi-3 for yourself, commencement with playing with the exemplary connected Azure AI Playground. Learn much astir gathering with and customizing Phi-3 for your scenarios utilizing the Azure AI Studio.

---

Footnotes

¹Table 1: Phi-3-small with lone 7B parameters

²Table 2: Phi-3-medium with 14B parameters

³Table 3: Phi-3-vision with 4.2B parameters

The station New models added to the Phi-3 family, disposable connected Microsoft Azure appeared archetypal connected Microsoft Azure Blog.