What Does Kimi K3 Actually Cost? Frontier Model Pricing Compared

2026年7月31日

Kimi K3 is an open-weight model released in July 2026. The weights are distributed, so in theory you can run it on your own machine. When I actually looked into it, though, nothing I have at home came close. So if you want to use this model, what does it cost?

This article lines up Kimi K3’s pricing against other frontier models (each vendor’s top-tier offering), and follows the process of actually signing up for it.

Checked on 28 July 2026. Prices move, so please confirm on the official pages before you commit to anything.

Can Kimi K3 run at home? Starting with the basics

Kimi K3 is a 2.8-trillion-parameter model. Running it locally means quantizing it (coarsening the data to shrink the file). Here are the sizes of the quantized versions currently being distributed.

Even the smallest 1-bit build is over 370GB. A mini PC with a large pool of unified memory tops out around 90GB in practice, so you are short by more than a factor of four.

Capacity is not the only obstacle. llama.cpp, the software most people use to run models locally, does not support Kimi K3 yet. K3 routes work internally across 896 “experts", while llama.cpp caps that number at 512 — it stops as soon as it tries to load the file. This cannot be worked around on the file side, according to the discussion around it.

I have written before about forcing a 745B model to run. That was 202GB squeezed onto a 128GB machine — an overflow of about 1.6×. That trick does not stretch this far.

Open weights or not, running this at home is not realistic. That is the starting point. If you want to use it, you will be signing up for a service.

Is Kimi K3 expensive? Lining it up against the others

API pricing is quoted per million tokens. A token is a chunk of text — for English, roughly three-quarters of a word.

Here are the main models, sorted by input price.

Lined up like this, Kimi K3 sits in the upper-middle of the field. Roughly level with GPT-5.6 Terra, above Claude Sonnet 5, below Claude Opus 5 and GPT-5.6 Sol. Being open-weight does not make it a bargain.

K3 does have two pricing advantages. First, when you send the same content repeatedly, caching brings input down to $0.30 — one tenth of list price. Second, it handles a one-million-token context with no tiered surcharge for length. Many vendors raise the rate once your context grows, so for feeding in long documents the gap could matter.

Pay-as-you-go or subscription? Finding the crossover

Kimi also sells subscriptions. K3 is available on the $19/month tier and above.

Which is cheaper depends on volume. Assuming a fairly ordinary input-to-output ratio of about 5:1, the blended effective rate works out to roughly $5 per million tokens. That balances against a $19/month plan at somewhere around 3.8 million tokens.

Coding agents burn tokens on a different scale, because they loop through many exchanges automatically. Running 50 million tokens a month on metered pricing would come to about $250 — thirteen times a $19 plan. In that territory the subscription wins outright.

Subscriptions do come with usage limits. Kimi’s are described as a rolling cap every five hours plus a seven-day reset, so it is not unlimited. Anything beyond the cap continues at roughly API rates as an add-on charge. For heavy use, the real shape is a subscription with metered usage on top.

Can you actually sign up? Payment is the real wall

This is where I got stuck, and it will affect anyone outside China.

The API flow itself is simple: create an account, issue an API key, deposit at least one dollar. Your rate limit rises in steps as your cumulative deposits grow.

The problem is how you pay. The official guidance lists only WeChat Pay and Alipay QR payments for personal accounts. I could find no mention of credit cards, and the official documents do not confirm whether a card issued outside China will work.

There is also language implying that an account is expected to be tied to a phone number. With Chinese services it is not unusual to find overseas phone numbers and email addresses failing verification, so it is worth assuming you may stumble at the door.

Regional resellers are starting to appear

After I had written most of this, I found that a Japanese company had begun offering Kimi K3 domestically. AI& Inc. started serving it through its own inference service on 27 July 2026, announced on the 29th.

That removes the payment wall entirely for customers in Japan — you are contracting with a local company, not opening an account with a Chinese service. The company also states that inference is completed within the country, which matters if you would rather your data did not leave.

Their published rates are $3.00 input and $13.00 output per million tokens. Output below the $15.00 of the original was not what I expected. Three of them side by side:

On capability, the reseller cites a third-party measure: the Artificial Analysis Intelligence Index, placing Kimi K3 at 57 and Claude Opus 5 at 61.

Four points apart on the index, but $13.00 against $25.00 on output — close to double. Measured as what you pay for the same job, the price gap is the more visible one. The company claims cost reductions of up to 80%; that is their figure, not something I have tested.

The point worth taking away is general: if the direct route is blocked where you live, a regional reseller may already exist, and it may not be more expensive.

The simpler workaround: go through an aggregator

OpenRouter, which fronts many models behind one account, offers Kimi K3 at the same $3.00 / $15.00. It takes credit cards, so you never contract with the Chinese service directly. For simply trying the model, this is the practical route.

Ollama can also reach a cloud K3, but that needs a paid plan from $20/month and bills by GPU time rather than tokens. Since you cannot compare unit prices, it is a poor fit if you are trying to forecast spend.

Another route: rent a GPU by the hour

Since the weights are public, you could rent GPUs by the hour and serve it yourself. But at over 370GB even in its smallest form, Kimi K3 means renting several datacenter GPUs at once — not something you try casually.

Where this approach earns its keep is trying a model in the tens-of-GB range before buying hardware for it. You can run something that would not fit on your machine, for a few dollars an hour, and find out whether it actually suits your work before you buy a graphics card for it.

One such service is RunPod.

The link above is a referral link; signing up through it credits both sides.

So what does fit at home? A rough guide

K3 is out of reach, but it is worth writing down what is not.

On a mini PC with a large memory pool, a 235B-class model ran at a genuinely usable speed. I wrote that up in Running huge LLMs on consumer hardware.

As an Amazon Associate we earn from qualifying purchases.

Summary

Kimi K3 is open-weight, but weights being published and the model being runnable at home turned out to be two different things. Cut down to one bit it still exceeds 370GB, and the software does not support it yet regardless.

Using it means paying for a service, and the price sits upper-middle among frontier models — open weights do not translate into a discount. Its pricing strengths are the flat rate across a very long context, and caching that drops input to a tenth.

The wall for anyone outside China is that official top-ups are limited to Chinese payment apps. An aggregator like OpenRouter sidesteps that with a credit card, and regional resellers are beginning to appear — one in Japan is already quoting output below the original. Casual trials point to the aggregator; business use points to a local provider.

Pricing in this field moves quickly. The figures here were checked on 28 July 2026, and I would confirm on the official pages before committing.

Postscript: people are still trying to run it locally

While writing this, the person distributing the quantized builds posted an update: fixing the 2-bit version and releasing a 1-bit version at the same time.

They also noted that running it in llama.cpp still needs a pending pull request to be merged, so it still does not run today. The fix for the expert-count ceiling described above has not landed upstream.

Put the other way round: once that merges, and if the size comes down to something like 100GB, a large-memory mini PC starts to look possible. I will write it up if that happens.

Note that these quantized builds are converted by individuals and are not official releases. Factor that in if you try them.

References

Hardware mentioned