We are excited to announce main updates to Azure AI Foundry, our built-in platform for designing, customizing, and managing enterprise-grade AI purposes.
At Microsoft, we’re devoted to advancing AI innovation to empower organizations, remodel industries, and redefine productiveness. Today, we’re excited to announce main updates to Azure AI Foundry, our built-in platform for designing, customizing, and managing enterprise-grade AI purposes. These updates embody groundbreaking new fashions like OpenAI’s GPT-4.5, enhanced fine-tuning and distillation methods, and the launch of latest enterprise instruments for brokers. These developments are designed to speed up the journey from AI experimentation to tangible enterprise influence.
Introducing GPT-4.5 in preview on Azure OpenAI Service
Building on the success of earlier fashions, GPT-4.5 is the newest and strongest general-purpose mannequin. This analysis preview demonstrates the achievements from scaling pre and post-training, a step ahead in unsupervised studying methods.
- Natural interplay: GPT-4.5 presents a extra pure interplay expertise. It has a broader information base, and its greater “EQ” might help to enhance coding, writing, and problem-solving duties.
- Accuracy and hallucinations: With a decrease hallucination charge (37.1% versus 61.8%) and better accuracy (62.5% versus 38.2%) in comparison with GPT-4o, builders can depend on extra exact and related responses.
- Stronger human alignment: Enhanced alignment methods enhance GPT-4.5’s means to comply with directions, perceive nuances, and have interaction in pure conversations, making it a more practical device for coding and challenge administration.
Developers can leverage GPT-4.5 in quite a few methods to reinforce productiveness and creativity. In communication, customers can depend on GPT-4.5 to craft clear and efficient emails, messages, and documentation. It additionally presents customized studying and training experiences, serving to customers to amass new expertise or deepen information in particular areas. During brainstorming periods, GPT-4.5 can generate revolutionary concepts and options, making it a precious device for inventive considering.
For challenge planning and execution, GPT-4.5 assists customers in organizing their duties, guaranteeing thorough and environment friendly approaches. It may also deal with complicated job automation, simplifying intricate processes and workflows. Developers can streamline their coding workflows by getting step-by-step steering and automating repetitive duties, to avoid wasting time and scale back errors. Overall, GPT-4.5 is a flexible mannequin. Starting at this time, enterprise prospects can entry GPT-4.5 in Azure AI Foundry. GPT-4.5, can also be obtainable in GitHub Copilot Chat to Copilot Enterprise customers.
New fashions: Phi-4, Stability AI, and up to date releases
The newest wave of AI fashions shares a standard focus: delivering specialised capabilities with higher effectivity. These releases signify a shift towards purpose-built AI that excels in particular domains whereas requiring fewer computational sources. Here are some standout launches on Azure AI Foundry.
Microsoft’s Phi fashions proceed to push the boundaries of what’s doable with smaller, extra environment friendly architectures:
- Phi-4-multimodal unifies textual content, speech, and imaginative and prescient for context-aware interactions. Retail kiosks can now diagnose product points by way of digicam and voice inputs, eliminating the necessity for complicated guide descriptions.
- Phi-4-mini packs spectacular efficiency into simply 3.8 billion parameters with a 128K-token context window. It outperforms bigger fashions on coding and math duties whereas rising inference velocity by 30% in comparison with earlier fashions.
Stability AI continues to advance generative imaging with fashions that speed up inventive workflows:
- Stable Diffusion 3.5 Large generates high-fidelity advertising and marketing property sooner than earlier variations, sustaining model consistency throughout various visible types.
- Stable Image Ultra achieves photorealism for product imagery, decreasing photoshoot prices by means of correct materials rendering and shade constancy.
- Stable Image Core an enhanced model of SDXL (text-to-image generative AI mannequin developed by Stability AI), gives high-quality output with distinctive velocity and effectivity.
Cohere enhances data retrieval capabilities with its newest rating expertise:
- Cohere ReRank v3.5 delivers extra correct search outcomes by means of its 4,096-token context window and assist for over 100 languages, surfacing related content material even with out precise key phrase matches.
The GPT-4o household expands with two specialised variants:
- GPT-4o-Audio-Preview handles audio prompts and generates spoken responses with applicable emotion and emphasis, which is good for digital assistants and customer support.
- GPT-4o-Realtime-Preview eliminates dialog lag with breakthrough latency discount, creating genuinely human-like interplay flows.
Agora, a pioneer in enabling real-time engagement, has been utilizing GPT-4o-Realtime.
GPT-4o-Realtime has revolutionized voice interplay for our conversational AI product, empowering builders with multilingual human-like voices, steady streaming, and ultra-low latency throughout customer support, telemedicine, and training.
—Patrick Ferriter, Senior Vice President, Product & Marketing, Agora
These advances collectively sign AI’s evolution towards extra pure, responsive, and environment friendly interactions throughout various use circumstances and deployment environments.
As our mannequin library surpasses 1,800 choices, we proceed to push the boundaries of experimentation and observability. Our new suite of fine-tuning instruments enhances the rise of unsupervised studying methods.
- Reinforcement fine-tuning: Now in non-public preview, this system teaches fashions to cause in new methods by rewarding appropriate logical paths and penalizing incorrect reasoning.
- Provisioned Deployment for fine-tuning: Azure OpenAI Service now presents Provisioned Deployments for fine-tuned fashions, guaranteeing predictable efficiency and prices by means of Provisioned Throughput Units (PTUs) along with token-based billing.
- Fine-tuning for Mistral Models: Exclusive to Azure AI Foundry, Mistral Large 2411 and Ministral 3B now assist fine-tuning for industry-specific duties like healthcare doc redaction.
Secure automation and scale for enterprise brokers
In at this time’s enterprise panorama, safety, and scalability are strategic imperatives. We’re introducing two highly effective options designed that can assist you securely harness AI for mission-critical duties:
- Bring your Vnet: Azure AI Agent Service now allows all AI agent interactions, information processing, and API calls to stay securely inside your group’s personal digital community, eliminating publicity to the general public web. Early adopters like Fujitsu are leveraging this functionality to enhance gross sales by 67% with their gross sales proposal creation agent, saving numerous hours that may be redirected towards buyer engagement and strategic planning, all whereas sustaining information integrity.
To study extra about constructing multi-agent purposes with Azure AI Foundry, try this informative webinar.
Create your AI resolution with Azure AI Foundry
We are enthusiastic about these new developments and sit up for seeing how you’ll leverage these highly effective instruments and options to drive innovation. Get going with new fashions and instruments in Azure AI Foundry.