Agents at Scale: Inside MiniMax's Model and the Infrastructure Behind It
AI Engineer World's Fair 2026 · 20:14
AI cloud infrastructure and model inference
Together AI runs a cloud platform for training, customizing and serving open AI models. Application developers can call models on demand through Together Inference, submit batches for asynchronous processing or deploy a model on dedicated hardware. With managed fine-tuning, a team can adapt a model to its own data, then serve it on Together or download the weights. Teams running their own training jobs can rent dedicated GPU clusters through Accelerated Compute. The platform also provides storage for model weights and datasets, and code sandboxes for applications and agents.
Vipul Ved Prakash, Ce Zhang, Percy Liang and Christopher Ré founded Together AI in 2022. The cost of the GPU clusters needed to train large models had concentrated that work in a few companies. The founders wanted other teams to have access to both training infrastructure and training data. One early collaboration, RedPajama, released an open dataset for training language models. Vipul serves as CEO and Ce as CTO. Together’s founding account
Before joining Together as chief scientist in July 2023, Tri Dao developed FlashAttention with Daniel Y. Fu, Stefano Ermon, Atri Rudra and Christopher Ré. The algorithm computes exact attention in blocks that fit in fast on-chip memory. This saves the GPU from repeatedly reading and writing a large intermediate matrix, a particular expense on long sequences. Tri’s follow-up, FlashAttention-2, split the work on a single attention head across multiple GPU thread blocks so more of the chip could work in parallel. Together made it available through its fine-tuning API and now also names Tri as a founder. FlashAttention-2 release · Together Research
Decagon used Together’s inference engine for customer-service voice agents, where a slow model response leaves a caller waiting in silence. The teams trained small models to predict the next few tokens of a reply so the larger model could check several predictions at once. Alongside prompt caching and other changes to how the models ran, this helped cut Decagon’s 95th-percentile model latency per turn from seconds to under 400 milliseconds, even on inputs tens of thousands of tokens long. Decagon case study
In July 2026, Together reported more than one million developers and thousands of paying customers, including Decagon, Cursor and Cognition. That month it announced an $800 million Series C at an $8.3 billion post-money valuation to expand its products and computing capacity. Series C announcement
AI Engineer World's Fair 2026 · 20:14
AI Engineer World's Fair 2026 · 30:00
AI Engineer World's Fair 2026 · 16:55
AI Engineer Europe 2026 · 24:35
AI Engineer Europe 2026 · 15:50
Loading more talks and speakers…
Affiliations reflect their AIE event listings and recorded appearances, not necessarily current employment.