MoonshotAI: Kimi K3

Kimi K3 is a multimodal reasoning model from Moonshot AI, released by the provider on 16 July 2026. It is an open-weight model with 2.8 trillion parameters, and it sits in the text category on api14 while accepting more than text as input. Moonshot AI positions it for complex coding, knowledge work, and long-horizon agentic workflows.

On the input side, Kimi K3 accepts text, images, and video, and it produces text as output. The context window is 1,048,576 tokens, which is large enough to hold very long documents, extended conversation histories, or sizable codebases in a single request. The maximum output is 943,718 tokens, so responses can themselves be extremely long when a task calls for it. Reasoning is part of the model's design, and the API exposes controls for it, including a reasoning parameter, a reasoning_effort setting, and an include_reasoning flag for surfacing the reasoning content.

Those capabilities point to a few practical uses. Long-context work such as reviewing large repositories, summarizing or querying lengthy reports, and maintaining state across multi-step agent runs fits the context window and the agentic focus. Because images and video can be passed in alongside text, the model can handle tasks that mix visual material with written instructions, for example describing or extracting information from screenshots, diagrams, or clips. Structured outputs, tool calling, and reasoning controls make it a reasonable fit for pipelines where the model must call functions, return JSON, or spend more effort on hard problems.

You call Kimi K3 through api14's OpenAI-compatible API using the model id "moonshotai/kimi-k3". The supported request parameters are frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, reasoning_effort, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs, and top_p. Sampling controls (temperature, top_p, top_k, min_p, the penalty parameters, seed, and stop) let you tune output, while response_format and structured_outputs cover formatted replies and tools and tool_choice cover function calling. Logprobs and top_logprobs are available if you need token-level probabilities, and max_tokens caps the response length within the maximum output.

There is no information here about pricing, latency, or throughput, so those are worth checking against the live price table and your own testing. The model is text-output only, so if you need generated images, audio, or other output modalities, a different model will be required. If your workload is small, short-context, and does not benefit from reasoning or multimodal input, a simpler model may be a better match.