The AI Stack
Sign in

Together AI

Cloud platform providing GPU infrastructure and APIs to run and fine-tune open-source AI models.

Updated May 2026

Overview

Founded
2022
Headquarters
San Francisco, CA
Ownership
Private
Segment
Model Distribution & Serving

Product overview

Together AI provides a cloud platform for running, fine-tuning, and deploying open-source LLMs and other AI models at scale with optimized inference and cost-effective GPU clusters. Enterprises, developers, and AI startups use it as a cost-effective alternative to proprietary model APIs for production workloads. Distinct for low-cost open-source model inference, custom fine-tuning pipelines, and large-scale GPU cloud capacity focused specifically on generative AI.

Revenue model

Usage-based API pricing per token (e.g., Llama 3.3 70B $0.18/M input, $0.59/M output; Qwen 2.5 72B $0.12/M input, $0.18/M output); custom fine-tuning billed per GPU-hour; enterprise plans with volume discounts and dedicated support.

Moat

Together AI's key competitive moat is its proprietary software stack and cutting-edge systems research, including custom CUDA kernels, Transformer-optimized kernels, quality-preserving quantization, speculative decoding, and the Together Kernel Collection, which deliver up to 2x faster inference, 60% lower costs, and 90% faster pre-training on GPU clusters compared to standard infrastructure. This performance optimization, combined with a full-stack AI platform enabling seamless training, fine-tuning, and deployment of open-source models on user-owned data with abstracted orchestration, creates high switching costs for customers reliant on its superior speed, economics, and developer tools.

Stack lineage

What powers Together AI

L4Models
  • DeepSeek

    Together AI hosts DeepSeek's models — e.g. deepseek-ai/DeepSeek-V4-Pro appears in their public model catalog.

  • Google DeepMind

    Together AI hosts Google DeepMind's models — e.g. google/gemma-4-31B-it appears in their public model catalog.

  • Stability AI

    Together AI hosts Stability AI's models — e.g. stabilityai/stable-diffusion-xl-base-1.0 appears in their public model catalog.

  • Allen Institute for AI (Ai2)

    Together AI hosts Allen Institute for AI (Ai2)'s models — e.g. allenai/Molmo-7B-D-0924 appears in their public model catalog.

  • OpenAI

    Together AI hosts OpenAI's models — e.g. openai/gpt-oss-120b appears in their public model catalog.

  • Meta AI

    Together AI hosts Meta AI's models — e.g. meta-llama/Llama-3.3-70B-Instruct-Turbo appears in their public model catalog.

  • Mistral AI

    Together AI hosts Mistral AI's models — e.g. mistralai/Voxtral-Mini-3B-2507 appears in their public model catalog.

L2Cloud
  • CoreWeave

    Together relies on large-scale GPU cloud capacity similar to GPU cloud providers

  • Alibaba Cloud

    Together AI hosts Alibaba Cloud's models — e.g. Qwen/Qwen3.5-397B-A17B appears in their public model catalog.

L1Silicon
  • NVIDIA

    Together AI hosts NVIDIA's models — e.g. nvidia/parakeet-tdt-0.6b-v3 appears in their public model catalog.

What Together AI powers

L6Applications
  • Pika

    Demi Guo, CEO of Pika: Together GPU Clusters provided training performance, expert support, and ability to scale for their AI creator community.

  • Krea

    Victor Perez, Co-Founder of Krea: Together AI provides performance and reliability for real-time image and video generation at scale.

  • Cursor

    Cursor partnered with Together AI to deliver real-time, low-latency inference at scale.

  • Decagon

    Decagon engineered sub-second voice AI with Together AI, achieving 6x cost reduction per turn and 11x faster inference.

L4Models
  • Cartesia

    Cartesia runs real-time voice AI on Together AI's GPU infrastructure.

  • Arcee AI

    Arcee AI moved from AWS to Together Dedicated Endpoints for greater inference flexibility, achieving 95% faster TTFT.