Together AI and Moonshot AI partner to serve Kimi K3 and future open models

Together AI and Moonshot AI announced a strategic partnership that makes Together AI the launch platform for Moonshot's open-weight model releases. The first model under this agreement is Kimi K3, a 2.8-trillion-parameter sparse Mixture-of-Experts model with native vision support and a 1-million-token context window. Together AI commits to serving every future open-weights Moonshot model at launch, with US-hosted infrastructure, zero data retention, and compliant access to models and data.

Model architecture and performance

Moonshot built Kimi K3 around two new architectural components. Kimi Delta Attention (KDA) changes how information flows across sequence length and delivers faster decoding at long context lengths. Attention Residuals (AttnRes) improves how representations are retrieved across model depth, adding training efficiency at minimal extra compute cost. Moonshot also applied optimizer refinements, including per-head Muon and quantile-based expert load balancing. According to Moonshot, these combine for roughly 2.5x better scaling efficiency compared to Kimi K2. The model targets long-horizon coding, agentic workflows, game development, and knowledge-intensive tasks, with benchmark results that Moonshot says put it in direct competition with leading proprietary systems.

Inference and post-training options

Kimi models are available across Together AI's inference products: Serverless, Provisioned Throughput, and Dedicated Inference. Provisioned Throughput offers reserved capacity with token-based pricing and a 99% uptime SLA, giving teams a predictable price without GPU-hour math. Dedicated Model Inference provides fast deployment and better token economics for teams wanting more control. Developers can also post-train Kimi models using custom training, including full-weight and LoRA reinforcement learning and advanced supervised fine-tuning. Training is configured through the Python SDK and granular primitives, and multiple LoRA experiments can run concurrently on dedicated capacity. Checkpoints deploy natively to inference without a separate handoff between training and serving.

What developers get

Day-zero availability means Kimi K3 is on Together AI with model quality validated by Moonshot, and future Moonshot releases will ship on Together at launch. Together AI's inference stack is tuned for large sparse MoE models, and it already serves production traffic for companies like Cursor, Y Combinator, and Decagon. As Moonshot ships new models, they land in the same Models library behind the same API, so developers avoid new SDKs, separate accounts, or re-plumbing their apps. Together customers can use Kimi models as a base for fine-tuning and can remove the attribution required in the MIT license. Because Kimi K3 is open-weight, developers can start serverless and move to Provisioned Throughput or Dedicated Inference as usage scales, without being locked into a single provider's roadmap.