Key research and product announcements at the AI Native Conf

At the first AI Native Conf, Together Research didn't just show up with a keynote — it showed up with seven separate ships in one day: FlashAttention-4, a Reinforcement Learning API, ThunderAgent, ATLAS-2, together.compile, a Megakernel implementation, and cache-aware prefill-decode disaggregation. That's a lot to unpack, but the real story isn't any single release. It's that the people writing the papers are the same people running the servers for customers like Cursor and Decagon, which means a technique doesn't sit in a repo for six months before it touches production.

Start with the kernels. FlashAttention-4, built under chief scientist Tri Dao, is tuned specifically for NVIDIA Blackwell and comes in 2.7x faster than Triton and 1.3x faster than cuDNN 9.13 — numbers that matter a lot once you're running long-context video or coding-agent workloads where every millisecond of attention compute gets billed. Then there's the Megakernel story, which is the kind of case study that actually explains why any of this matters: a real-time voice agent company needed sub-100ms time-to-first-64-tokens and was stuck at 281ms on B200 GPUs. Together's team rebuilt the whole model into a single hand-tuned kernel targeting H100 memory bandwidth and got them to 77ms — a 3.6x speedup and 7.2x better economics. together.compile, meanwhile, tries to automate that kind of tuning instead of requiring a specialist for every model; on Hedra's video model it cut generation time 25%, and on Flux Kontext benchmarks it beat torch.compile by 41%.

On the training side, Together's new Reinforcement Learning API is built around a blunt fact: over 70% of RL wall-clock time is spent on rollouts, which is just inference wearing a different hat. So the API exposes rollout configuration and weight-push frequency as knobs teams control, while Together handles synchronization — pushing new weights to inference nodes in seconds within a datacenter, under a minute globally. ThunderAgent, open-sourced today, tackles a related and less obvious problem: agentic workloads like coding agents get treated as disconnected, stateless requests by most inference systems, which causes KV cache thrashing, lopsided GPU load, and leaked Docker sandboxes. Treating each agent run as one schedulable unit instead fixes all three, and Together claims 1.5–3.6x throughput gains for agentic serving and up to 3.9x for RL rollouts.

The two inference optimizations closing out the announcements go after subtler inefficiencies. ATLAS-2 turns speculative decoding from a static artifact into something that keeps learning — using accepted and rejected tokens from live traffic to retrain the draft model on the fly and hot-swap new versions in without downtime, adding roughly 1.2x on top of existing static speculators and, crucially, not decaying as traffic shifts. Cache-aware prefill-decode disaggregation solves a routing problem: cold 100K-token prompts and warm multi-turn requests were competing for the same prefill capacity, so CPD splits them across dedicated tiers with a three-level KV-cache hierarchy, delivering 35-40% higher sustainable throughput on B200s under coding-agent-style traffic.

Together frames all of this as a flywheel — research improves the platform, the platform attracts workloads, those workloads expose the next hard problem, repeat. It's a tidy story, and the seven releases at least back it up with actual before-and-after numbers rather than vibes.