The first models in the Llama 4 herd Available today in Azure AI Foundry and Azure Databricks, let users create more individualised multimodal experiences. Designed by Meta, these models are meant to smoothly combine text and vision tokens into a single model backbone. Setting a new standard in AI development, this creative solution lets programmers use Llama 4 models in applications needing large volumes of unlabelled text, image, and video data.
Top intelligence Unmatched speed and efficiency
Llama’s most accessible and scalable generation is here. Unmatched efficiency, step performance changes, extremely long context windows, mixture-of-experts models, native multimodality. All in simple-to-deploy sizes tailored fit for your intended usage.
Most recent designs
The models are optimised for simple deployment, cost-effectiveness, and performance scaling to billions of users.
Llama 4 Scout
Meta claims Llama 4 Scout, which fits in a single H100 GPU, is more powerful than its Llama 3 models and among the best multimodal models in its class. It also raises the supported context length from 128K in Llama 3 to an industry-leading 10 million tokens. This creates a whole new world of options, such multi-document summarization, parsing large user activity for tailored activities, and reasoning across large codebases.
Targeted applications include reasoning, personalization, and summarization. It excels in jobs needing compressing or analysis of large information because to its quick size and long context. It can create reports or summaries from very long inputs, tailor its responses using thorough user-specific data (without losing prior information), and conduct sophisticated reasoning across vast knowledge bases.
Llama 4 maverick
Low cost, fast replies with revolutionary intelligence make industry-leading natively multimodal model for image and text understanding. Offering good quality at a lesser cost than Llama 3.3 70B, Llama 4 Maverick is a general-purpose LLM with 17 billion active parameters, 128 experts, and 400 billion total parameters. With support for 12 languages, Maverick shines in image and text comprehension, hence allowing the development of complex artificial intelligence systems spanning linguistic boundaries. Perfect for exact image understanding and creative writing, Maverick is quite appropriate for general assistant and chat applications. It provides cutting-edge intelligence with quick speed, optimised for optimal response quality and tone for developers.
Targeted use cases include those for conversation that have been optimized and need for top-notch replies. Meta fine-tuned 4 Maverick to be a great conversational agent. The main conversation model of the Meta Llama 4 family think of it as the multilingual, multimodal equivalent to a ChatGPT-like assistant.
It works especially well with interactive applications:
- Bots that provide customer service must comprehend the photos that consumers upload.
- AI creative collaborators with multilingual communication and content creation skills.
- Internal enterprise assistants who can assist staff members by managing rich media input and responding to enquiries.
Businesses may use Maverick to create superior AI assistants that can communicate with a worldwide user base in a natural (and courteous) way and use visual context when necessary.
Llama 4 Behemoth Sneak Peek
A still-in-training early glimpse of the Llama 4 instructor model employed to distil Scout and Maverick.
Features
Class-leading features are provided by Llama 4 Behemoth, Scout and Maverick.
Naturally Multimodal: Designed with natural multimodality, all Llama 4 models use early fusion to pre-train the model using significant quantities of unlabelled text and vision tokens, a step change in intelligence from distinct, frozen multimodal weights.
Unmatched Long Context: Supporting up to 10M tokens of context, Llama 4 Scout opens up new possibilities for memory, personalization, and multi-modal applications by providing the greatest context length now available in the sector.
Image Grounding Expert: Best-in-class on image grounding, Llama 4 can match user requests with pertinent visual ideas and tie model reactions to areas in the picture.
Writing in many languages: Supporting worldwide development and deployment, Llama 4 was also pre-trained and fine-tuned for unmatched text comprehension across 12 languages.
Benchmark
Testing for coding, reasoning, knowledge, vision understanding, multilinguality, and lengthy context, Meta assessed model performance on a set of common benchmarks spanning a wide spectrum of languages.
