Moonshot’s Kimi K3 Makes Video a Native Input in a 2.8-Trillion-Parameter Open-Weights Model

Moonshot AI’s Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model with a 1-million-token context window and native multimodal input. The important detail for production teams is not only that it can inspect images. Kimi K3 can accept video alongside text and images, letting a prompt address the footage itself instead of a transcript or a manually selected set of frames.

Kimi K3 treats video as vision content inside a multimodal prompt. A team uploads a clip, references it in the request, and pairs it with a text instruction such as identifying a visual event, checking continuity, or locating a moment that matches a brief. Moonshot says it will release K3’s weights on July 27, 2026, opening a self-hosting and fine-tuning path for teams that need more control over their material.

Moonshot’s vision-model documentation describes multimodal requests as a sequence of text and visual content. For video, a team uploads the clip and references it in the request through a file ID or video_url, alongside the instruction. The model samples key frames from that video and can return text about the visual material it sees.

That distinction matters in a media workflow. A transcript can tell a model what was said. A contact sheet can show a handful of selected moments. A video request gives K3 sampled key frames from the clip, so it can assess visual material across a sequence rather than only a manually selected still-image set. It is not a frame-by-frame reading of every moment, but it can connect a visual event, an action, or a screen graphic to the prompt without first reducing the entire clip to text.

The company also pairs that capability with a 1-million-token context window. In practice, the useful question is not whether a facility can fit an entire production in one prompt. It is whether a clip, its related notes, a transcript, and a detailed review instruction can be kept together without reducing the visual material to text before analysis.

Claude Opus can analyze image inputs, and its vision documentation describes image content blocks for that work. It does not document a native video-input block in the same way. A Claude workflow that needs to reason about footage therefore starts by converting the material into inputs the model accepts, such as extracted frames and a transcript.

Kimi K3 does not make those workflow artifacts obsolete. A transcript remains useful for dialogue search, and selected frames remain useful for precise visual review. Its difference is that the video file can be part of the model request, with the model sampling key frames rather than relying only on a preselected still-image set. That makes K3 a model to test for logging footage, reviewing a clip against a brief, or combining visual questions with production notes.

This is a capability comparison, not a blanket quality claim. The right model still depends on the task, the material, the evaluation criteria, and whether a team can run the model reliably on its own infrastructure. Moonshot's public materials identify video as an input modality, but production teams should test accuracy on their own footage before relying on it for editorial, compliance, or asset-management decisions.

Kimi K3 uses a mixture-of-experts architecture, which activates only part of its total parameter count for a given request. Moonshot says the model has 2.8 trillion parameters, 1 million tokens of context, always-on reasoning, and native multimodality. It also highlights Kimi Delta Attention, which it says improves decoding speed at long context lengths, and Attention Residuals, which it says improve training efficiency.

Those are Moonshot's performance claims, and they need independent testing. The immediate operational fact is that the model is available through Kimi's products and API, while the company says the full open weights will follow on July 27. If that release arrives as described, facilities will be able to assess a very large model with text, image, and video input without being limited to a hosted API.

For a studio or post house, self-hosting does not automatically mean simple or inexpensive deployment. A model at this scale will require substantial infrastructure, and any video-analysis workflow will need careful tests around clip preparation, latency, security, and failure cases. Open weights do, however, make it possible to evaluate, adapt, and operate the model within a team's own environment rather than sending every request to a third-party endpoint.

Before deploying K3, post teams should build a test set of representative clips, define the visual questions the model must answer, and compare its sampled-frame results with a transcript-and-contact-sheet workflow. They should also measure latency, infrastructure cost, and the rate of missed visual events before using the model for any consequential production decision.