Together AI announces strategic partnership with Moonshot AI to natively serve Kimi models

Together AI just locked down a deal that's less about one model and more about owning a pipeline. Moonshot AI, the lab behind the Kimi series, will ship every future open-weights model on Together AI first, starting now with Kimi K3, a 2.8-trillion-parameter sparse mixture-of-experts model that's the largest open model released to date.

K3 isn't just big for the sake of headlines. It packs a 1 million token context window and native vision support, and Moonshot built two new architectural pieces into it: Kimi Delta Attention, which speeds up decoding at long context lengths, and Attention Residuals, which improves how the model retrieves information across its own depth. Add in tweaks like per-head Muon optimization and quantile-based expert load balancing, and Moonshot claims roughly 2.5x better scaling efficiency than its predecessor, Kimi K2. That's the kind of efficiency gain that actually matters when you're running a model this size in production.

The real story, though, is what Together AI is building around it. Kimi K3 lands across Serverless, Provisioned Throughput, and Dedicated Inference, meaning teams can start cheap and scale into SLA-backed capacity, 99% uptime, without doing GPU-hour math themselves. There's also full post-training support baked in — LoRA and full-weight reinforcement learning, supervised fine-tuning, and a Python SDK — so a fine-tuned checkpoint can go straight to serving without a separate handoff. Together says its stack already handles agentic and coding workloads at scale for customers like Cursor and Decagon, which is presumably why Moonshot picked this partner over rolling its own hosting.

What's notable is the standing commitment: every open Moonshot model, not just K3, ships on Together AI on day zero, through the same API and account, no re-plumbing required each release. For developers who've been burned by chasing model updates across five different providers, that consistency is worth more than the benchmark numbers.