MoonshotAI: Kimi K3 (batch)

MoonshotAI: Kimi K3 (batch) is a text generation model from Moonshot AI, identified on api14 by the model id moonshotai/kimi-k3:batch. It is a multimodal reasoning model with open weights and 2.8 trillion parameters, built for demanding work rather than lightweight chat. The batch variant is the version you call through api14, and it keeps the same model identity and behaviour as the base Kimi K3 release from the provider.

The model accepts text, images and video as input, and it produces text as output. Its context window is 1,048,576 tokens, which is large enough to hold entire codebases, long document sets or extended conversation histories in a single request. The maximum output length is 16,384 tokens per response. Because it is a reasoning model, it can spend part of its generation budget on internal reasoning before producing the final answer, and that behaviour can be influenced through the reasoning and reasoning_effort parameters.

Those capabilities point to a few practical uses. Complex coding tasks fit well, including working across large repositories where the full source needs to be in context. Knowledge work such as summarising, comparing or extracting from long documents and reports is another natural fit, especially when the source material includes images or video that must be interpreted alongside text. Long-horizon agentic workflows are also a target use case, where the model plans over many steps, calls tools and keeps track of state across a long context.

You call the model through api14's OpenAI-compatible API using the model id moonshotai/kimi-k3:batch. The request supports frequency_penalty, include_reasoning, logit_bias, max_tokens, min_p, presence_penalty, reasoning, reasoning_effort, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k and top_p. Tool calling and structured outputs are available through the tools, tool_choice, response_format and structured_outputs parameters, and sampling can be tuned with temperature, top_p, top_k, min_p and the penalty parameters. A seed can be set for reproducible sampling.

Keep the output ceiling in mind: responses are capped at 16,384 tokens, so tasks that need a very long single answer may require splitting the work across calls. The model is also text-output only, so any image or video input is interpreted to produce text rather than a generated media output. For simple, short text tasks, a smaller and faster model may be a more economical choice than a 2.8T parameter reasoning model.