Moonshot AI has launched Kimi K3, a model with 2.8 trillion parameters, a 1-million-token context window, and built-in multimodal support, designed for long-horizon coding and complex knowledge work (Moonshot AI). The model is available through Moonshot's API platform, with pricing details at platform.moonshot.ai/docs/pricing/chat-k3 (Moonshot AI Platform).

The 2.8-trillion parameter count places Kimi K3 among the largest deployed AI models today. Parameters are the internal weights a model uses to make predictions; more parameters generally mean a more capable model, though the relationship isn't perfectly linear. The more operationally meaningful number is the context window: 1 million tokens. A token is roughly three-quarters of a word, so the model can accept about 750,000 words of input in a single call. That is enough for entire codebases, long legal documents, or multi-file research projects, all processed at once rather than split into smaller pieces (Moonshot AI; Moonshot AI Platform).

Native multimodality means Kimi K3 handles text and other input types (such as images) within a single model rather than passing them through separate specialized systems. For developers building workflows that combine document understanding, code generation, and data extraction, this removes the overhead of stitching together multiple models and the latency that introduces (Moonshot AI).

Moonshot is positioning Kimi K3 specifically for long-horizon work. The word "long-horizon" matters here: this model is not tuned for short back-and-forth chat. It targets multi-step reasoning over large volumes of context, where the model must stay coherent across tens of thousands of tokens of intermediate steps, tool calls, and partial results before delivering a final output (Moonshot AI Platform).

On the commercial side, Moonshot is running a promotion: pre-funding an account on the Kimi API Platform earns users up to 30% back in vouchers (Moonshot AI Platform). This kind of rebate is a standard customer-acquisition tactic among API platforms competing for developer attention. It lowers the effective per-token cost for early adopters willing to prepay, which subsidizes experimentation and migration at a stage when developers are still evaluating whether Kimi K3's quality justifies the integration effort.

The pricing page at platform.moonshot.ai/docs/pricing/chat-k3 is the authoritative source for per-token costs, though the specific rates should be checked live, as API pricing can change independently of model announcements (Moonshot AI Platform).

The broader context for developers and enterprises evaluating Kimi K3 comes down to three factors: context length, parameter scale, and multimodal integration. The 1-million-token window reduces or eliminates the need for retrieval-augmented generation (RAG) in many workflows. RAG is a technique where external systems fetch relevant text chunks to feed into a model because the model's context window is too small to hold everything at once. Cutting that scaffolding can lower infrastructure complexity and remove failure modes that chunking introduces. The 2.8-trillion parameter count raises questions about inference economics: larger models require more compute per token, and the cost per query depends heavily on Moonshot's serving infrastructure, batching efficiency, and whether the architecture uses techniques like mixture-of-experts to activate only a fraction of total parameters during each forward pass. The publicly available facts do not specify the architecture in that level of detail.

Native multimodality simplifies the technology stack but does not guarantee that the model matches specialized single-modality systems on every input type. Developers integrating Kimi K3 should benchmark it against existing systems on their specific workloads rather than treating parameter counts or context-window specs as reliable proxies for real-world performance.

For teams already on Moonshot's platform, the voucher promotion lowers the cost of stress-testing K3 against production workloads. For those considering a switch, the decision hinges on whether the combination of extended context and multimodal handling offsets switching costs and the uncertainty that comes with any newly deployed model at this scale.

The available facts are limited to Moonshot's own disclosures and platform documentation. No independent benchmarks, third-party evaluations, or comparative performance data appear in the verified sources, so claims about Kimi K3's standing relative to competing models would be speculative. What can be reported is the model's specifications, its positioning, its availability through Moonshot's API, and the promotional pricing incentive currently in effect.