A more accurate model was waiting for $1.45.
We chose not to use it.

On July 27, 2026, Moonshot released the weights for Kimi K3, the newest model in the kimi ai line. That same day we reached for it through its API to read a 165-page Thai book whose text layer was broken end to end, not a single page the machine could read as characters.

Two days later a second book showed up, 85 pages, the exact same task. The same more accurate model was sitting right there. Reading the whole book would have cost $1.45. And we chose not to use it.

That decision is the heart of this review. Kimi K3 is a strong model, but "strong" turned out to be the wrong thing to grade it on. Here is what it actually is, what it costs, the license clause most people miss, and the question that made the choice easy.

What Kimi K3 actually is

Kimi is Moonshot AI's model line, the brain behind the kimi ai assistant. K3 is the latest, and the news that matters is that Moonshot released the actual weights. That's what "open weight" means: in principle you can stand the model up on your own hardware instead of being locked to one vendor's service. Hold onto that word "principle," because the specs decide how real it is.

Here are the four specs that change a decision, and only those:

• 2.78 trillion parameters, mixture-of-experts. Each token wakes up only 16 routed experts plus 2 shared ones, not the whole model. That's how a model this large stays affordable to run per query, but the full thing is still enormous.
• It ships 4-bit, and it's still 1.56TB. The download is 1.56TB across 118 files, and it's only that "small" because the released checkpoint is already quantized to 4-bit. This is server-class hardware, not a laptop. "Open weight" hands you the option to self-host, it does not make it free or easy.
• Multimodal. It reads images, not just text, which is the only reason we could point it at a PDF whose text layer was dead and have it read from the pictures.
• Roughly 1M-token context. It can hold a whole book in one pass, so you're not stitching chapters back together by hand.

One honest note for this review: we did not download those 1.56TB. We accessed K3 through Moonshot's hosted API, which is how most people will actually meet it. So take every number below as "using K3," not "running K3 ourselves."

What it costs to actually use it

Reading that first 165-page book through the hosted API cost about $0.017 per page, so roughly $2.4 for the whole book. Cheap, for what it did. But the useful part of a Kimi K3 review isn't the sticker price, it's the trap that isn't in any doc.

K3 can reason before it answers. Leave that "thinking" mode on while you're just transcribing text, and it burns tokens on reasoning until the budget runs out, then sometimes hands back an empty file while you've paid in full. Our first run that way ran nearly triple the cost per page for nothing. The fix: switch thinking off (with it off, the API pins the randomness setting to 0.6, and that's fine). One line of config saved most of the bill.

The wider lesson, which we've written about before in verifying what a model reads, is that the numbers that decide a job aren't on the pricing page. You only get them by paying for one real run.

The license catch nobody reads

People keep calling it "MIT" or "Modified MIT." It isn't. The model card lists license: other, and the file is titled "Kimi K3 License." It borrows MIT's wording, then adds four conditions MIT never had. Two of them matter if you plan to build on it:

• A model-as-a-service clause: if you turn K3 into a service others call, and that business grows past a set revenue line, you need a separate agreement with Moonshot.
• An attribution requirement: past 100 million monthly users or $20 million a month in revenue, you have to show the words "Kimi K3" on screen.

Here's the part worth the read. Clause 4 carves out an exemption: internal use that you don't expose to outsiders, and access through Moonshot's official API or a certified partner, are free of both conditions entirely. In plain terms, the lane most people will actually use, calling the hosted API for internal work, slips out of the scary clauses cleanly. That's the story, not the word "MIT."

The question that makes this useful to you

Choosing a model used to be a straight line between accurate and cheap. Open weights add a third axis: the model can come to the data, instead of the data having to go to the model. Once the weights exist in the open, in principle you can stand them up wherever the data lives. That reframes the whole choice.

So the first question isn't "how hard is this task," it's can this data leave home. Which is exactly why we said no to K3 on the second book. It had real people's names in it, and a chapter on getting consent to use someone's voice. A voice is biometric data. The moment it goes into an API, it leaves home into the hands of someone who has never seen it. That settled the question before price or accuracy entered the picture.

And refusing the more accurate model cost us nothing that mattered. We sent that book to a cheaper model already inside our plan, no new data leaving, and measured what broke. The Thai prose came out mangled, but the code and the config values came out correct, and the code was the one thing we actually wanted. Picking the "worse" model wasn't giving up. It was getting clear on what we needed.

Verdict: who Kimi K3 is for

Kimi K3 earns the hype on capability: multimodal, huge context, and genuinely open weights with a license that gets out of your way for internal and API use. As a review verdict, it's a real frontier model you can build on.

But grade it on the right thing. Reach for K3, or any open weight model, when your data is the constraint, when the win is that the model can come to the data. If your data can leave home freely, a hosted closed model may still beat it on ease and raw accuracy, and that's a fine answer too. If you're weighing self-hosting more broadly, we've mapped that in fitting a local LLM to the job.

The question was never how smart Kimi K3 is. It's whether your data has a passport. Answer that first, and the model almost picks itself.