Together AI

Together AI positions itself as an "AI native cloud": a full-stack platform covering inference, accelerated compute and model shaping for teams building AI applications. Its inference products include serverless inference for running open-source models on demand, batch inference for asynchronous processing of large workloads (stated to scale to 30 billion tokens per model with any serverless model or private deployment), provisioned throughput with reserved capacity, token-based pricing and a stated 99% uptime SLA plus drop-in API compatibility, dedicated model inference on isolated infrastructure, and dedicated container inference aimed at generative media workloads such as video, audio and image models.

On the compute side, the company offers accelerated compute ranging from self-serve instant clusters to thousands of GPUs, optimized with the Together Kernel Collection, including on-demand B200 availability on Together GPU Clusters. Adjacent infrastructure services include code sandboxes for building development environments for AI apps and agents, and managed storage combining object storage and parallel filesystems for AI workloads with no egress fees. Model shaping is represented by fine-tuning of open-source models for production use, positioned as improving accuracy, reducing hallucinations and controlling model behavior without customers managing training infrastructure. The platform claims 2x faster inference, 60% lower cost through workload-specific optimization, and 90% faster pre-training via the Together Kernel Collection.

The company also runs a research organization, Together Research, that publishes work spanning inference optimization, GPU kernels, model architecture and agents — for example FlashAttention-4, Mamba-3, cache-aware prefill–decode disaggregation, consistency diffusion language models, speculative decoding for reinforcement-learning rollouts, and agent benchmarks and datasets such as ParallelKernelBench, DSGym, EinsteinArena and CoderForge-Preview. Papers are co-authored with researchers affiliated with universities including Princeton, CMU, Stanford-affiliated authors, UC Berkeley, Seoul National University and Georgia Tech, and Together AI hosts an AI Native Conf and presents work at ICML.

Business model

Together AI sells cloud infrastructure and platform services for AI workloads, spanning on-demand and dedicated inference, GPU clusters, sandboxes, storage and fine-tuning, with self-serve sign-up alongside a sales-led route for larger deployments.

Sources describe usage- and capacity-based offerings: serverless, on-demand inference with no long-term commitments; batch inference for asynchronous workloads; provisioned throughput sold as committed capacity with token-based pricing; and dedicated model or container deployments.

Latest developments

Recent announcements on the company site include a partnership with Y Combinator to deliver a dedicated YC GPU cluster, on-demand B200 GPUs on Together GPU Clusters, serving of MiniMax-M3 for efficient inference, and product and research announcements made at the AI Native Conf.