Kimi K3 is a new AI model from Moonshot AI, a Chinese AI lab, released on July 16, 2026. It has 2.8 trillion parameters, making it the largest open-weight model to come out of China, and on parts of the benchmark leaderboard it already beats OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8.
Here's the full picture: what Kimi K3 does well, how it compares to the models you already know, and how to try it yourself without writing a line of code.
The quick answer
Kimi K3 is Moonshot AI's flagship model: 2.8 trillion parameters, a 1-million-token context window, and native vision, meaning it can read images as well as text. Moonshot calls it the first "open 3T-class" model. It's available now through Kimi's website, its mobile app, and its API.
Full self-hosting isn't possible yet. The model's weights, the actual trained parameters, aren't public. Moonshot has committed to releasing them by July 27, 2026, which depending on when you're reading this may have already happened.
Why Kimi K3 is all over your feed right now
K3 lands in the middle of a run of open-weight releases out of China. DeepSeek, Z.ai's GLM series, and Moonshot's own earlier Kimi K2.6 have each closed more of the gap with what OpenAI and Anthropic charge a premium for. K3 is the biggest step yet. On Moonshot's own benchmark results, K3 beats Claude Opus 4.8 and GPT-5.5 across most of its evaluation suite, though it still trails Anthropic's top model, Claude Fable 5, and OpenAI's GPT-5.6 Sol.
Independent testing backs that up. Developer and AI researcher Simon Willison reported that on Artificial Analysis's private long-horizon knowledge-work evaluation, K3 scored 732 Elo points higher than Moonshot's previous model, Kimi K2.6, ranking second only to Claude Fable 5. K3 also took the top spot on Arena.ai's Frontend Code leaderboard, ahead of Fable 5.
The timing adds to the story. TechCrunch reported, citing the Financial Times, that Moonshot was raising a funding round that would value the company at $31.5 billion, up from a $20 billion valuation in May. The release also landed inside an active industry argument about whether companies should keep paying for closed models like Claude and GPT-5.6, or shift toward cheaper open-weight models and train them for their own use. K3 gives the "cheaper and open" side of that argument a much stronger example to point to.
What "open-weight" means (and why K3 isn't fully open yet)
Open-weight means Moonshot publishes the trained model's parameters, the billions of numbers that encode what it learned, so anyone with enough hardware can download and run it themselves. Closed models like GPT-5.6 and Claude only run on their maker's servers. Open-weight isn't the same as fully open-source software either, since Moonshot isn't necessarily releasing the training code or the data K3 learned from alongside the weights.
The weights themselves aren't public yet. Moonshot has committed to publishing them by July 27, 2026, but as of this post, K3 is only reachable through Kimi's own website, app, and paid API. For a home user, that barely matters. You can already open the chat and use it. The company running a coding agent or a fine-tuning pipeline is the one waiting on the weights.
How Kimi K3 compares to the models you already know
Moonshot's own release notes are unusually direct about where K3 falls short. It lists "a noticeable gap in user experience" compared to Claude Fable 5 and GPT-5.6 Sol as a known limitation, alongside a tendency toward "excessive proactiveness," where the model makes decisions on the user's behalf during long, ambiguous tasks.
Pricing sits closer to the top end than you'd expect from an open-weight model. K3 costs $3 per million input tokens and $15 per million output tokens (less if a request hits Moonshot's cache), which puts it in the same range as Anthropic's Claude Sonnet series and makes it the most expensive model a Chinese lab has released to date. That's a real jump from Kimi K2.6's $0.95/$4 pricing. Artificial Analysis estimated K3's typical cost per task at around $0.94, close to GPT-5.6 Sol's $1.04 and roughly half of Opus 4.8's $1.80.
One quirk worth knowing before you try it: K3 launched with a single thinking-effort setting, "max." Willison ran his usual test prompt, asking the model to draw an SVG of a pelican riding a bicycle, and it cost 25 cents because the model used more than 13,000 reasoning tokens on a request that should take a fraction of that. Moonshot says lower-effort modes are coming in a later update.
How to try Kimi K3 yourself
You don't need to be a developer to use it. Go to kimi.com or download the Kimi app, available on iOS, Android, and HarmonyOS, and sign in with a Google account or a phone number. No credit card required for the free tier.
One thing to check: the chat may default to Moonshot's older K2.6 model. Open the model picker and switch it to K3 before you start. The free tier covers chatting with K3, uploading files, and web search, though the full 1-million-token context window and higher thinking-effort settings are reserved for paid tiers.
What this means for you
For most people reading this, Kimi K3 is a signal about where the AI market is headed: a frontier-level model that costs less to build on than the top US labs' offerings, and one that's about to be open enough for other companies to build products on top of it. If you're a developer or a small team weighing whether to keep paying for a closed model or experiment with an open one, K3 is worth a real look once the weights land. If you're a curious reader who just wants to see what the fuss is about, kimi.com gets you there in under a minute.
If you want the same breakdown for the models you're probably already using, see what's inside ChatGPT's current lineup, Claude's model picker, and Gemini's model lineup.
Sources:
kimi.com/blog/kimi-k3 ·
simonwillison.net/2026/Jul/16/kimi-k3 ·
techcrunch.com, "Moonshot's upcoming Kimi 3..." ·
cnbc.com, "China's Moonshot AI unveils Kimi K3..."
