Vipul Ved Prakash, Co-Founder and Chief Executive Officer of Together AI, established the enterprise alongside Ce Zhang and Percy Liang.
The founders built the organisation on the premise that open models would eventually close the quality gap with proprietary systems, believing businesses would ultimately seek more adaptable and affordable tools to deploy them.
Vipul says: “Our mission is to ensure that intelligence is abundant, not expensive.” This strategic bet has paid off as models from MiniMax, DeepSeek, Kimi and NVIDIA have steadily advanced open-weight capabilities towards the technical frontier.
Supporting inference performance
To cater to these evolving workloads, the platform supports the entire lifecycle of an AI model.
Software engineers can access hosted models using an API, fine-tune systems on proprietary company data or lease dedicated GPU clusters for large-scale training workloads.
Through this comprehensive offering, Together AI now supports more than one million developers alongside thousands of paying commercial accounts.
Behind the scenes, its internal research team focuses heavily on inference speed, recognising that low latency is critical for running voice agents, code assistants and other real-time software.
Practical deployment data demonstrates the financial and operational benefits of this performance-led approach. For instance, Decagon reduced its cost per turn sixfold compared with a closed alternative while operating sub-second voice agents.
Similarly, Cursor runs real-time inference with low latency on systems powered by 72 NVIDIA GB200 NVL72 GPUs.
Salesforce AI Research cut system latency by 50% and lowered operational costs by approximately one-third, while Zomato doubled customer satisfaction scores after deploying a support bot capable of managing more than 1,000 messages each minute.
Leave a Reply