Chat with Kimi K3.

Moonshot’s largest model, at 2.8 trillion parameters. It reads text, pictures and video, and is aimed at hard coding and at work that runs for a long time.

Start chatting with Kimi K3 now — free, no account

At a glance

  • Maker: Moonshot AI
  • Released: 2026
  • Parameters: 2.8T
  • Context window: 1M
  • Size: Very large
  • Memory: Handles very long documents
  • Good at: Hard coding and long agent runs
  • Supports: tools, images, zero data retention
  • License: Modified MIT
  • Upstream API price: $8.50 per million tokens

What is Kimi K3?

Kimi K3 is the largest model on UsingOpen: 2.8 trillion parameters from Moonshot AI, the Beijing lab behind the Kimi assistant. Moonshot made its name on long context before it was fashionable, and K3 represents the other thing the lab is known for — building at genuinely frontier scale and releasing the weights. At 2.8T it out-muscles every other model listed here.

The specification reads like a wish list: a one-million-token context window — roughly 750,000 words — text, picture and video input, tool support for agentic work, and zero-data-retention terms. The build is aimed at two things in particular: hard coding and long agent runs. The price is the honest counterweight: $8.50 per million tokens, the highest on the site, which makes K3 a model you reach for when the problem is worth it rather than a daily driver.

Moonshot's philosophy with the Kimi line has been more context, more scale, more care, and K3 is the current end of that sentence. If a smaller model has almost solved your problem, K3 is what you hand it to. If you are price-sensitive, the fast tier will do most of this work at a fraction of the cost — the art is knowing which problems are which.

What can you do with Kimi K3?

K3 exists for the top of the difficulty curve:

  • Hard coding: gnarly concurrency bugs, subtle architecture calls, performance work that resists analysis.
  • Very long agent runs — tasks measured in hundreds of turns and many hours.
  • Whole-project comprehension with everything in one window, roughly 750,000 words.
  • Questions about pictures and video answered at flagship quality.
  • Final-verification passes: checking another model's work with a bigger model.
  • Research synthesis across very many long documents.

How does Kimi K3 compare?

Its only true peer on the site is Alibaba's Qwen3.8 2.4T ($6.00) — the other open heavyweight, with 95B of its 2.4T parameters awake per token. K3 is bigger, reads video as well as pictures, and is aimed harder at coding; the Qwen is a little cheaper and plays the same heavyweight-reasoning game. Between them they set the ceiling of open-model spending here.

Below the heavies, DeepSeek's V4 Pro ($1.98) gets you most of the flagship experience at a quarter of K3's price, and GLM 5.3 ($2.43) is the coding-focused alternative. K3's premium buys the last increment of scale — worth it for the hardest jobs, extravagant for ordinary ones.

Is Kimi K3 right for you?

K3 is the model you pick when you have already tried something cheaper and it was not enough: the bug that survives three models, the agent run that needs to stay coherent for hours, the analysis where almost-right is a failure mode. At $8.50 per million tokens it is the most expensive model on UsingOpen, and the only honest defence of that price is the one the specification makes — it is the biggest, at 2.8 trillion parameters, and among the most capable, with text, pictures and video in, tools supported, and a million tokens of working memory.

Everything else about it is agreeable. The zero-data-retention terms are attached, the context is as long as context gets, and the coding focus is real rather than incidental. The question with K3 is never whether it can do the job; it is whether the job needs K3.

For daily work, it does not: the Flash tier covers ordinary coding and chat at a fraction of the price. For verification work — checking a smaller model's answer on something that matters — and for the genuinely hard residue, it does. Teams that keep a heavyweight on standby for escalations will find K3 the most capable standby on this site, with Qwen3.8 2.4T its only real competition at the top.

Frequently asked questions

Is Kimi K3 the biggest model on UsingOpen?

Yes — 2.8 trillion parameters, larger than every other model on the site, including Qwen3.8 2.4T and the DeepSeek V4 Pro releases. If the job calls for sheer scale, this is the ceiling. Second place is not close on price either — Qwen3.8 2.4T costs $6.00 and the DeepSeek V4 Pros under $2 — so K3 is unambiguously the ceiling of open spending here.

What is the context window of Kimi K3?

One million tokens — roughly 750,000 words. Moonshot built its reputation on long context, and K3 carries the full million alongside its frontier scale.

Is Kimi K3 worth $8.50 per million tokens?

It depends on the job. For hard coding, very long agent runs and problems where a smaller model gets you 90 percent of the way, the premium buys the last stretch. For ordinary drafting and Q&A, the fast tier does the same work at a fraction of the price. A practical discipline: track which of your tasks actually needed the escalation. Most teams find it is five to ten percent, which is exactly the shape K3 is for.

Does Kimi K3 support image input?

Yes. It reads text, pictures and video, so visual material joins the conversation natively alongside code and prose. Reading video as well as pictures at this scale is rare; it makes K3 usable for material that other coding flagships would refuse at the door.

Who makes Kimi K3?

Moonshot AI, the Beijing lab behind the Kimi assistant. It is known for pioneering very long context windows and for releasing large open-weight models under permissive MIT-style terms.

Is Kimi K3 good for coding?

Yes — hard coding and long agent runs are its two stated aims. This is the model for the bugs and designs that resist smaller models. Its reputation was earned on the ugly problems — the concurrency bug, the hot path, the refactor nobody wants — rather than on tidy examples.

Is Kimi K3 free to chat with on UsingOpen?

Yes. Pick it in the model picker and start chatting — no account, no payment. History is saved only on your device, and the picker shows live availability. Free access to the biggest open model on the site is worth testing even if you never expect to need it.

Related models on UsingOpen

Models most often compared with Kimi K3 on UsingOpen:

  • Qwen3.8 2.4T by Alibaba: the open heavyweight — 2.4T parameters with 95B awake for the hardest reasoning.
  • DeepSeek V4 Pro by DeepSeek: the biggest DeepSeek — the full-strength 1.7T V4, built for hard problems.
  • GLM 5.3 by Z.ai: the full-size flagship — 753B parameters for the longest, hardest coding jobs.
  • Inkling by Thinking Machines: the larger Thinking Machines model — 975B parameters that read text, pictures and sound.

Head-to-head comparisons

See how Kimi K3 lines up with the models people most often weigh it against:

Model availability changes as upstream capacity comes and goes — the picker on the chat page shows what is ready right now.

Chat with Kimi K3 free on UsingOpen — no account, no sign-up

← All open-weight models · Compare models · Free AI chat · FAQ