Together AI

2,048 reasons to train with us. Leading AI companies like Pika, Nexusflow, INSAIT - Institute for Computer Science, Artificial Intelligence and Technology train on Together Cloud. We have optimized the infrastructure and software for large scale training and inference. We just added 2,048 top spec H100s to our fleet in Together Cloud. Reserve yours today while supply lasts: https://lnkd.in/ghS_3NYb

RAG fine-tuning combines code retrieval & model training, addressing outdated knowledge & hallucinations in LLMs. Our experiments show RAG fine-tuned models using the Together AI Fine-tuning API and Morph Labs Code API achieve up to 16% better accuracy, 3.7x faster speed & 150x cost reduction vs Claude 3 Opus, and up to 19% quality improvement, 1.1x faster speed & 37.5x cost reduction vs GPT-4o. https://lnkd.in/digqW5Nk

We are honored to be named in Redpoint's 2024 InfraRed 100, celebrating the most transformative infrastructure companies. Our team is committed to helping developers and businesses build and run Generative AI applications with the best performance and cost at production scale. We're hiring for roles across our teams. If you think you can help us, please reach out! https://lnkd.in/gHmD7PAt

Vals.ai has built a third-party review system to evaluate AI model performance in different industries such as accounting, law, and finance. They needed an AI platform to run their eval suite on a variety of industries and across multiple benchmarks. By choosing Together AI, Vals.ai has been able to run ~ 320k API calls, 200M tokens in a single day on Together AI while keeping their costs low and steady. This has enabled them to test new models and add them to their leaderboard on the same day they're released. Read more about Vals.ai’s journey on Together AI here https://lnkd.in/gh9KvTDt

Mixture of Agents—a framework that leverages the collective strengths of multiple LLMs. Each layer contains multiple agents that refine responses using outputs from the preceding layer. Together MoA achieves a score of 65.1% on AlpacaEval 2.0. Together MoA uses six open source models as proposers and Qwen1.5-110B-Chat as the final aggregators with three layers. We also evaluate on FLASK which offers more fine-grained evaluation and outperforms original models on most of the dimensions. Both Together MoA and Together MoA-Lite are on the Pareto front, indicated by the dashed curve, in the performance vs. cost plot. Try Together MoA through our interactive demo. Please note that the TTFT is slow at the moment due to the iterative refinement process of MoA, but we are actively working on optimizations. 📣 Blog: https://lnkd.in/g26Ya_98 📄 Paper: https://lnkd.in/g_vc9Hac ⌨️ Code: https://lnkd.in/gF2xEEeb 🖥️ CLI Demo: https://gttps://https://lnkd.in/gifKtUQv This work was made possible through the collaborative efforts of several open-source projects. We appreciate Meta, Mistral AI, Microsoft, Alibaba Cloud, and Databricks for developing the Llama, Mixtral, WizardLM, Qwen, and DBRX models. We also thank Tatsu Labs, lmsys.org, and KAIST AI for the AlpacaEval, MT-Bench, and FLASK evaluation benchmarks.

We are excited to announce the launch of Dragonfly, an instruction-tuned vision-language architecture that enhances fine-grained visual understanding and reasoning about image regions! The Dragonfly architecture employs multi-resolution visual encoding and zoom-in patch selection to enhance multimodal reasoning while being context-efficient. We are also releasing two new open-source models: 1. Llama-3-8b-Dragonfly-v1: A general-domain model trained on 5.5 million image-instruction pairs, demonstrating promising performance on vision-language benchmarks like visual QA and image captioning. 2. Llama-3-8b-Dragonfly-Med-v1: A biomedical model fine-tuned on 1.4 million biomedical image-instruction pairs, in collaboration with James Zou's group at Stanford University School of Medicine. Dragonfly-Med outperforms prior models, including Med-Gemini, on multiple medical imaging tasks, showcasing its capabilities for high-resolution medical data. We hope this open-source release will benefit the research community in exploring multimodal AI applications and addressing real-world problems. Read more on our blog: https://lnkd.in/g3jDk34w Arxiv Paper: https://lnkd.in/gsbYk6mV Codebase for architecture implementation: https://lnkd.in/gsBk9n6P

At Together AI we have an increasing number of customers and partners leveraging the Together API for multi-agent workflows and applications. The ability to leverage multiple models, with industry-leading performance has made Together Inference a leading solution for agent-based workflows. Recently we worked with Axiomic, a new AI agent build framework, who integrated the Together API into the core framework as the provider for using open-source models. In our most recent blog we explore agent building using one of Axiomic's demos called GEAR Chat, which shows how four agents work together to provide a portable, steerable, and debuggable chat agent. Read more here: https://lnkd.in/gsBPviiG

Latitude.io has built a new era of video games where the player is not limited to what the developer has pre-imagined, but is dynamic. With Together AI, they were able to address challenges in hosting large models, optimizing GPU deployments, and managing AI development costs. Specifically, by leveraging the scalable and fast Together Inference, Latitude increased their total daily tokens by 8x while keeping latency low using Llama-3 and Mixtral. They were able to provide more coherent models for players by providing a 13B model to free players, and also achieve a 37x reduction in safety classifier costs. Read more about it here: https://lnkd.in/gh9KvTDt

We are thrilled to share that our founder and CEO, Vipul Ved Prakash, will be speaking on "Adoption of Open Source Models in Global Enterprises" at SuperAI in Singapore this Thursday, June 6th. Join us and industry leaders for an insightful discussion! If you're attending, don't miss the panel at 3:40 PM on the Main Stage. Add it to your calendar here: https://lnkd.in/gD95EPt5

Yesterday, Cartesia released Sonic, a blazing fast lifelike generative voice model and API. Sonic is built on a new state space model architecture developed for efficiently modeling high resolution signals like audio and video. We are thrilled to partner with and enable Cartesia to serve their state-of-the-art custom state space model in less than 2 weeks and achieve the fastest text-to-voice generation in the market with 135 ms model latency to provide real-time inference to their users. Cartesia’s next-generation modeling capabilities coupled with Together AI’s deep understanding of the inference stack have enabled more flexible, low-latency serving of this text-to-voice model. In partnering with Together AI, Cartesia was able to configure and optimize for real-time inference with low latency at high accuracy, and low cost at production scale. Read more about our Solutions: https://lnkd.in/gh9KvTDt  Try Sonic: https://play.cartesia.ai