Together AI 与 Moonshot AI 达成战略合作，原生托管 Kimi 模型

Today we're announcing a strategic partnership with Moonshot AI, one of the leading model labs pushing the open source frontier with large-scale MoE architectures. Under this partnership, Together AI becomes a launch platform for Moonshot's model releases, starting with Kimi K3 and extending to every open weights model Moonshot ships going forward.

For developers building on open models, this means day zero access to Moonshot's frontier releases through Together AI’s US-hosted infrastructure, pricing model, and tooling, while ensuring zero data retention and compliant access to models and data. Developers can also post train these models with their own data to deliver the quality and performance for their app.

Kimi K3 is the largest open model released to date: a 2.8T parameter sparse Mixture-of-Experts model with native vision support and a 1M token context window. Moonshot built it around two new architectural components:

Layered on top are a set of optimizer refinements, including per-head Muon and quantile-based expert load balancing, that Moonshot says combine to deliver roughly 2.5x better scaling efficiency compared to Kimi K2. The result is a model built for long-horizon coding, agentic workflows, game development, and knowledge-intensive tasks, with benchmark results that put it in direct competition with the leading proprietary systems on the market today.

Kimi models are available across Together AI Inference products – including Serverless, Provisioned Throughput and Dedicated Inference. Developers can go from trying these models to applying them to their production use cases with SLA-backed products and autoscaling.

With Provisioned Throughpu t, teams can use reserved inference capacity for frontier open models with token-based pricing and a 99% uptime SLA. It’s the reserved-capacity guarantee developers already expect from closed-model providers, now available for open weights. No GPU-hour math, just guaranteed throughput at a predictable price. For teams looking for more control with all the benefits of a production-ready platform, they can use Dedicated Model Inference with fast deployment, better token economics and continuous research innovations shipped into the product.

Developers can also post train Kimi models to deliver better quality for their target use case. With custom training, including full-weight and LoRA reinforcement learning as well as advanced supervised fine-tuning, developers can configure training through the Python SDK and granular primitives, and run multiple LoRA experiments concurrently on dedicated capacity. Custom training connects experimentation directly to production. When a checkpoint is ready to evaluate, it can be deployed natively to inference, with no separate handoff between training and serving and no rebuilding.

Want to hear directly from the team behind Moonshot AI and Together AI? Join Feihu Tang from the Moonshot AI team and Jue Wang and Zain Hasan from Together AI to talk through the technical decisions behind Kimi K3, how to actually run it on Together, and what's coming next. Bring your questions.

RSVP and set a reminder for the Moonshot x Together Webinar