Together AI

Together AI's Decagon case study reports p95 model latency reduced from seconds to below 400 milliseconds and nearly sixfold lower cost per turn, using fine-tuned models, custom speculative decoding and prompt caching in production voice workloads.

Together AI reports thousands of customers across AI-native companies and operates a full-stack platform for accelerated compute, training, fine-tuning and production inference.

Together AI reports thousands of customers across AI-native companies and operates a full-stack platform for accelerated compute, training, fine-tuning and production inference.

Together published Tri Dao's FlashAttention-2 work with open-source implementation, sequence-length parallelism and sliced-Q work partitioning that reduces communication between GPU warps.

Together AI had nine papers accepted at ICML 2026 and conducts research across agents, inference, kernels, model shaping and AI systems, with research deployed into its production platform.

Together AI had nine papers accepted at ICML 2026 and conducts research across agents, inference, kernels, model shaping and AI systems, with research deployed into its production platform.

Together AI reports that it built a dedicated GPU cluster for Y Combinator startups, with self-service resource reservation, provisioning and billing, and describes the cluster as fully utilized. The operating customer-specific infrastructure is distinct from the existing financing and general platform-portfolio observation.

Together AI raised an $800M Series C in 2026 and offers production-scale compute, serverless and dedicated inference, fine-tuning, sandboxes and managed storage for AI workloads.

Together AI raised an $800M Series C in 2026 and offers production-scale compute, serverless and dedicated inference, fine-tuning, sandboxes and managed storage for AI workloads.

Together's ThunderKittens article describes a transparent CUDA/HIP embedded DSL organized around matrix tiles, tensor-core use and a PyTorch-like API. These concrete kernel-programming abstractions support an analyst inference of portable GPU optimization skills; the article is not evidence that every employee uses them.

Together AI hires across research and engineering while working from GPU kernels through model shaping and large-scale agentic inference, giving employees highly portable full-stack AI systems experience.

Together AI hires across research and engineering while working from GPU kernels through model shaping and large-scale agentic inference, giving employees highly portable full-stack AI systems experience.