Together AI announced seven research and product releases at AI Native Conf spanning GPU kernel optimization, reinforcement learning infrastructure, and inference efficiency. FlashAttention-4 targets NVIDIA Blackwell GPUs and is 2.7x faster than Triton. Together Megakernel runs entire models in a single kernel, achieving 3.6x performance gains for real-time voice agents. together.compile automates kernel optimization with a single function call, delivering up to 41% speedups over torch.compile. The Reinforcement Learning API exposes inference and training as separate configurable layers, with ThunderAgent adding program-aware scheduling for agentic workloads (1.5–3.6x throughput gains). ATLAS-2 introduces an online speculative decoding flywheel that continuously adapts to live traffic. Cache-aware prefill-decode disaggregation (CPD) adds a three-tier KV cache hierarchy, improving sustainable throughput by 35–40% for long-context workloads.
