Three inputs, one conversation

Watch how people actually work with an AI assistant and you will see a lot of window switching. The screenshot lives in one app, the voice memo in another, the written brief in a third, and the model sees none of the connections between them. The idea arrives fragmented, so the answer comes back fragmented too.

Kimi K3.1 Pro is built against that fragmentation. Kimi K3.1 Pro is an independent playground and API service for Moonshot AI's flagship model, and there text, images, video and transcribed audio land in the same thread, with the model reasoning across all of it at once. This review looks closely at how those input types work, because the input design is where the product differentiates itself.

The setup takes two minutes

Before examining the mechanics, the practical part: creating an account takes a Google sign-in, and new accounts receive 10 welcome credits. There is no card required to explore the Kimi K3.1 Pro playground. Email verification-code sign-in is currently disabled, so Google is the door in. Payment only enters the picture when the welcome credits run out and you decide Kimi K3.1 Pro has earned a place in your workflow.

That low-friction start matters for a multimodal product specifically, because you cannot judge how a model handles your sketches and recordings from a marketing page. You have to feed it your own material.

Text: the anchor input

Text remains the anchor. A brief, a difficult question or a rough draft goes into the thread, and Kimi K3.1 Pro responds with an explanation, a rewrite or a structured plan. Nothing surprising here, and none is claimed. What makes the text input interesting is everything that can sit next to it in the same message.

Consider the example the product itself demonstrates: a product sketch attached as an image, a voice note attached as audio, and a written instruction that says, in effect, look at the sketch, listen to my notes, and turn both into a launch plan. Kimi K3.1 Pro pulls the key decisions from two media types and merges them into one practical next step. No single-modality workflow can do that without the user manually bridging the gap.

Images: show, don't describe

Image input is where the model's native vision does the heavy lifting. PNG, JPEG and WebP files are supported, and the use cases divide into three recognizable groups.

First, description: hand over a screenshot or photograph and ask Kimi K3.1 Pro what it shows, down to details that are tedious to type out. Second, comparison: put two designs or two charts side by side and ask for the differences that matter. Third, explanation: show a diagram or an interface and ask why it is built the way it is.

For designers and product people, the comparison case is the quietly valuable one. Reviewing a wireframe against an earlier revision is exactly the kind of visual judgment that text-only pipelines force you to fake with written descriptions. Here the actual pixels are in the thread, and the judgment is made on them.

Audio: transcribed first, then reasoned

The audio path is worth describing precisely, because the service is transparent about how it works. Upload a recording and Kimi K3.1 Pro transcribes the speech first; reasoning then runs over the transcript alongside your prompt. The upstream model currently lists native text, image and video inputs, so audio arrives through speech-to-text preprocessing rather than raw waveform understanding.

That design has a practical upside even beyond simplicity: transcripts live in the thread, so you can see exactly what the model heard, and transcription cost is included in the normal usage billing rather than appearing as a surprise line item. For spoken briefs, meeting fragments and quick thoughts recorded on a walk, the two-step path is imperceptible in practice.

Video and the long thread

Video input extends the same pattern to recorded material. Alongside it sits the property that makes all the modalities genuinely useful together: the upstream model behind Kimi K3.1 Pro carries a 1,048,576-token context window. A million tokens is enough that a long document, a set of diagrams and an hour of transcripts can stay in one conversation instead of being chopped across sessions.

The site is honest about the boundary: service request limits also apply, and uploads are capped at 4 MiB per file, 8 MiB total and 8 files per conversation. Those caps shape the workflow toward focused sessions rather than dump-everything archives, which is a reasonable trade for a service that settles billing on actual usage.

The thread is the product

Here is the framework worth taking away: the individual input types are table stakes, and the real feature is that they share one thread with shared context. A question about the sketch can reference the voice note. A follow-up about the transcript can pull in a diagram. Context accumulates inside Kimi K3.1 Pro instead of being re-explained.

That accumulation is also what makes the API half of Kimi K3.1 Pro coherent. The same conversation shapes are available through an OpenAI-compatible endpoint, so a multimodal flow prototyped in the playground moves into an application without redesign. Same account, same keys, same credit balance.

Who should pay attention

Product managers who think in sketches and voice memos. Researchers who need documents and charts in one place. Developers who want a frontier model with vision and a 1M-token window behind a standard API. And anyone tired of duct-taping three apps together to assemble one thought.

Ten welcome credits are enough to test all of it with real material, and Kimi K3.1 Pro routes Moonshot AI's flagship through OpenRouter rather than pretending to be an official Moonshot property, a candor its own FAQ leads with. That kind of disclosure makes a third-party tool easier to trust. If your ideas arrive in more than one medium, the case for trying it writes itself: one thread, every input, and a model that reads all of it together.