Together AI

🚀 Announcing the launch of Llama 3.2 and Llama Stack on Together AI, in partnership with AI at Meta. 🎉 We are excited to offer free access to the Llama 3.2 vision model for developers to build and innovate with open source AI. Start building with the Llama-Vision-Free model today: 👉 https://lnkd.in/gWxwdaVd ▶ What we are launching: - Free Llama 3.2 Vision Model (11B): Develop and experiment with our high-quality Llama-Vision-Free endpoint for multimodal tasks. - Together Turbo Inference Endpoints (11B, 90B): High performance and accuracy for tasks like image captioning, visual question answering, and image-text retrieval. - New Llama Stack APIs: Standardized APIs to simplify building agentic and retrieval-augmented generation (RAG) conversational apps. ▶ Unlock powerful use cases: - Interactive Agents: Build AI agents that process both image and text inputs Image Captioning: Create high-quality image descriptions for e-commerce and digital accessibility - Visual Search: Enable users to search via images, enhancing search efficiency for retail and e-commerce ▶ Industry applications: - Healthcare: Accelerate medical image analysis for faster diagnostics Retail & E-commerce: Revolutionize shopping with image-based search and personalized recommendations - Finance & Legal: Streamline workflows by analyzing visual and textual content to optimize contract reviews and audits 💡 Check out napkins.dev: Our open-source demo app uses Llama 3.2 vision to transform sketches and wireframes into React code! Try it out at https://napkins.dev 🔧 Get started today: Experiment with the Llama-Vision-Free endpoint, or build for production with Llama 3.2 Together Turbo endpoints. 🌟 Read more in the blog https://lnkd.in/gDfmXxzu

👨🍳 New Cookbook: A step-by-step walkthrough of the PDF to Podcast workflow! In short it involves: 1. Structured decoding with Llama-3.1-70B on Together AI to extract a JSON script. 2. Cartesia TTS to bring the script to life! Notebook: https://lnkd.in/g7vS_4nV

We’re excited to power the AI infrastructure of leading enterprises like Salesforce, The Washington Post, and Zoom. With the Together Enterprise Platform, you can securely run inference, fine-tuning, and training on your own models, in your own environment, while unlocking 2-3x faster inference speeds and reducing operational costs by up to 50%. Learn more about our enterprise offerings at our new page: together.ai/enterprise

Announcing Together Cookbooks! 👨🍳 A collection of hands-on notebooks showcasing powerful use cases of open-source models with Together AI, including Text RAG, Multimodal Document RAG, Semantic Search, Rerankers, and Structured JSON extraction. Here's a glimpse at what’s inside:   🖼 Multimodal Document RAG with Nvidia Investor Slide Deck: Implement multimodal RAG with ColQwen2 and Llama 3.2 90b Vision, combining text and images for advanced retrieval. 📊 Embedding Visualization: Visualize vector embeddings to explore structure in high-dimensional spaces. 🌐 Knowledge Graphs with Structured Outputs: Generate knowledge graphs from LLMs using structured JSON generation. 🧠 Semantic Search: Boost search precision with BERT-based embedding models for better retrieval. 🔎 Improving Search with Rerankers: Refine search results with rerankers to enhance relevance across large document corpora. 📰 Structured Text Extraction from Images: Extract structured text from images, ideal for document digitization and workflow automation. 📝 Text RAG: Implement text-based Retrieval-Augmented Generation to enrich responses with relevant knowledge. 🔗 Explore the cookbooks here: https://lnkd.in/g5E4VYqd

Our CEO, Vipul Ved Prakash, joined Clara Shih, CEO of Salesforce AI, on the Ask More of AI podcast to dive into the future of generative AI and how Together AI is leading the charge. In this episode, they break down how Together AI is optimizing AI workloads to make models faster, smarter, and more efficient for real-world applications. Key innovations like FlashAttention, speculative decoding, and model quantization are highlighted, showing how we’re transforming AI workloads for greater speed, scalability, and impact. It’s an exciting look at how we’re helping businesses bring AI from pilot to production at scale. 🎧 Tune in to the conversation here: https://lnkd.in/g8HU3P4B

📬 Never miss a beat! Today we are introducing "Together We Build", a newsletter with a handpicked selection of news, product launches, novel research, and AI tools from Together AI. Subscribe to keep up with the latest developments in generative AI and LLMs! And don't miss our first issue 👇

New work on linearizing LLMs! Like subquadratic capabilities? Like modern 7B+ LLMs? But don’t have the budget to pre-train billions of parameters on trillions of tokens to get subquadratic, 7B+ LLMs? Then check out LoLCATs, our new work led by Michael Zhang that converts existing Transformers like Llama and Mistral into state-of-the-art subquadratic variants. Now for the same cost as a LoRA finetune! LoLCATs builds on a simple framework to convert Transformers into subquadratic models: 1. Swap an LLM’s softmax attentions for more efficient alternatives. 2. Fine-tune the LLM to adapt to these layers & recover pre-trained quality. However, to improve linearized LLM quality, while drastically reducing the cost of this recovery, we build LoLCATs around two simple findings. First, we can learn how to approximate softmax attentions with existing linear attentions. This lets us replace softmax attentions with near-literal drop-in replacements, that are still subquadratic to compute. Next, this makes parameter-efficient fine-tuning like LoRA, sufficient to adjust for any approximation errors & rapidly recover LM quality. The results speak for themselves. LoLCATs-linearized Llama 3 8Bs and Mistral 7Bs significantly outperform both prior linearized LLMs and strong subquadratic LLMs, while only training 0.2% of their parameters on 0.003 - 0.02% of their training tokens. And we did one last thing! Mostly just because we could, we used LoLCATs to linearize the entire Llama 3.1 family. And deliver the first linearized 70B and 405B LLMs, while significantly improving over baseline qualities. Learn more on our blog here: https://lnkd.in/gQFeZxMY

New Blog post: How to build a realtime AI image generator with Flux Schnell and Together AI! We discuss: - Generating & refining images with Flux Schnell - Debouncing API requests to make them "realtime" - How we built "consistency mode" in 1 line of code Read here: https://lnkd.in/eApuAbAq

### In this event we'll discuss how you can perform RAG over complex PDF documents that contain images, graphs, tables text charts and more! We'll describe in detail how: - The new image retriever ColPali works - How you can finetune ColPali to improve further for your usecase - How to leverage multi-vector retrieval to retrieve from PDFs - How to use language vision models like the new Llama 3.2 vision series to perform document RAG

Congratulations to Braintrust on their Series A! 🎉 We've been partnering with them to showcase the performance of our models in real-world tests. Check out this Braintrust recipe showing that Llama 3.2 vision models running on Together AI are 3x faster with the same accuracy as GPT-4o-mini and GPT-4o. Read here: https://lnkd.in/gq-jDNWq