Guide · checked 2026-10-20 · kimi-k3

How to call Kimi K3
2.8T · 1M · thinking will not turn off

As of 2026-10-20: Kimi K3 shipped on 16 Jul 2026, model id kimi-k3, official context window 1,048,576 tokens (1M). Official price per million tokens: $3 input miss, $0.30 cache hit, $15 output; cache write is its own line item, $3 for TTL 5min and $6 for TTL 1h, and a hit is billed at $0.30 with no repeated write charge. Thinking mode cannot be switched off; reasoning_effort has low / high / max, default max. QCode /models lists kimi-k3.

Updated 2026-10-22

Claude, GPT and Chinese models on one key, billed per token; check /models for callable models.

#kimi-k3#2.8T#1M#$3 / $0.30 / $15

Four numbers before you call

2.8T

Scale

Official: 2.8T, native vision, KDA + Attention Residuals. They call it the first open 3T-class model, and still say it trails Fable 5 and GPT-5.6 Sol — that is their own positioning.

1M

Context

Official 1,000,000-token window (the price table writes 1,048,576 tokens). The docs state that billing is not tiered by context length: everything is billed by usage, and no length surcharge appears on the official table.

$3 / $0.30 / $15

Official API / million

Cache-miss $3, cache-hit $0.30, output $15. From platform.kimi.ai, same band across the full 1M window. The price table checked on 2026-10-20 lists cache write as its own item: $3 for TTL 5min and $6 for TTL 1h; a hit is billed at the hit price only, with no repeated write charge.

Always on

You cannot disable CoT

Official FAQ: You can’t — K3 always thinks. Do not pass a thinking parameter. reasoning_effort is low / high / max, default max.

What K3 is

Moonshot's flagship thinking model. The official docs say: Kimi K3 is Kimi's most capable flagship model to date, with 2.8 trillion parameters, native visual understanding and a 1-million-token context window; it is the world's first open 3-trillion-parameter model, built for long-horizon coding, knowledge work and reasoning. The launch default was max thinking; later docs added low / high / max, but thinking itself still cannot be switched off. The full model weights are released (they had been promised for late July), and architecture, training and evaluation details follow with the Kimi K3 technical report.

Do not mix this with the other two pages

/kimi-k3-status-tracker tracks what shipped and what is confirmed; /kimi-k3-vs-claude-opus-5 is the item-by-item matchup with Opus 5. Price, cache write, thinking level and context window follow the official pages; third-party resale rates are not the official table.

Timeline

16 Jul 2026

Kimi K3 launches on kimi.com / Kimi Work / Kimi Code / platform.kimi.ai. Id kimi-k3.

late Jul 2026

Official promise back then: full weights by late July. The official model page checked on 2026-10-20 now states that the full model weights have been released, and the Hugging Face model card says the weights come under the Kimi K3 License; architecture, training and evaluation details follow with the Kimi K3 technical report.

This page

How-to. QCode /models lists kimi-k3, billed by token, with per-model prices on /models. Official price table checked on 2026-10-20: cache write is now its own line item.

Confirmed vs treat carefully

Confirmed

Ship date 16 Jul 2026; 2.8T; 1M context (the official price table writes 1,048,576 tokens); native vision; id kimi-k3; official $3 / $0.30 / $15; thinking always on; effort low / high / max, default max. The price table checked on 2026-10-20 also lists: cache write by TTL tier, $3 for 5min and $6 for 1h, with a hit billed at $0.30 only and no write charge on a hit; the official model page states that the full model weights have been released. Sources: the Kimi K3 tech blog, plus the model page and pricing page of the Kimi API platform.

Treat carefully

What Moonshot has not published yet: the full technical report (architecture, training and evaluation details) is announced to follow with the Kimi K3 technical report; vision input does not support public image URLs, so you need base64 or an uploaded file id; web search is being updated and the docs advise against using it in production workflows for now. Do not treat a reseller rate as Moonshot’s tariff. Moonshot itself says it still trails Fable 5 and GPT-5.6 Sol.

Hosted API or self-host

Call the API

model=kimi-k3. Official price $3 / $0.30 / $15. Per million tokens, with cache write billed separately by TTL. Thinking cannot be disabled; only reasoning_effort moves. QCode lists this id.

Want the weights

Released by now. The promise was late July; the Hugging Face model card says the full weights come under the Kimi K3 License, and the official model page checked on 2026-10-20 states the weights are released. Architecture, training and evaluation details follow with the Kimi K3 technical report.

How to call it

Pick kimi-k3 on the official platform. Do not pass thinking; it is not how you turn reasoning off. To shorten traces, set reasoning_effort to low. The docs require it: multi-turn conversations and tool calls must return the complete assistant message (including reasoning_content) unchanged. Moonshot says coding workloads often see cache-hit rates above 90%. Checked on 2026-10-20: max_completion_tokens defaults to 131072 and can be set up to 1048576; temperature, top_p and the rest are fixed, so leave them out.

On QCode

QCode can call kimi-k3, and /models already lists this id. Billing: by token, with per-model prices on /models. Model availability follows the live list on /models. Not ready for a monthly plan? Pay-as-you-go Packs start at $4 for a one-time quota you use within 7 days. If you already have a Personal Plan and run out of quota, use a booster pack.

Kimi K3 how-to FAQ

What is the model id?

kimi-k3. That is the string on the official API platform and on QCode /models.

Can I turn chain-of-thought off?

No. Official text: You can’t — K3 always thinks. If it runs long, set reasoning_effort to low. Do not pass a thinking parameter.

What is the official price?

Per million tokens: $3 for a cache miss, $0.30 for a cache hit and $15 for output, the same band across the whole 1M window, from platform.kimi.ai. The price table checked on 2026-10-20 lists cache write on its own: $3 for TTL 5min and $6 for TTL 1h; a hit is billed at $0.30 only, with no repeated write charge.

Are the weights downloadable?

Released. They were promised for late July; the Hugging Face model card now says the full weights come under the Kimi K3 License, and the official model page checked on 2026-10-20 states the weights are released; architecture, training and evaluation details follow with the Kimi K3 technical report.

Can I call it on QCode?

Yes. /models lists kimi-k3. Billing is by token, with per-model prices on /models; availability follows the live list on /models.

What is the status of Kimi K3 right now — is it still callable?

Shipped and callable: Kimi K3 launched 2026-07-16, the official model page states the full model weights are released, and QCode /models lists kimi-k3. The tracker is confirmation status, the vs page is Opus 5; specs follow the official pages. What is still missing: the full technical report and the evaluation details.

Sources

The Kimi K3 tech blog (kimi.com/blog/kimi-k3); the Kimi API platform pages for the Kimi K3 model, model inference pricing, context caching and thinking models; the English model page and pricing page on platform.kimi.ai. All fetched and checked on 2026-10-20.

The id is kimi-k3

It is listed. Thinking stays on. Per-model prices on /models.

Not sure where to start?

Try it for a week

Small Pack: $4 for $30 of quota, valid 7 days

Use it every day

Starter: $9.00/month with $5 of quota every day — just $0.060 per $1 of quota, vs about $0.113 for the Medium Pack