Free Kimi API access: what is actually free.

Quick answer: four routes, four catches.

Self-hosting is the fifth route and gets its own section below, because for this family the hardware math is the whole story. The decision between the routes reduces to what you need to stay working: if a capped mirror going away mid-project costs you nothing, use it; if you are evaluating the real models for real work, evaluate on the ids and rate behavior you would ship with.

• Moonshot's platform (platform.moonshot.ai): no standing free API tier. Third-party guides checked in July 2026 report small trial allowances for new accounts and a first-recharge bonus, and at least one reports that account activation requires a minimum top-up. Terms have shifted more than once, so verify in the console before planning around them.
• The consumer Kimi app: genuinely free chat usage on the free plan, but it is a chatbot, not an API key. Searches for a free Kimi API often end here by mistake.
• Community free mirrors: aggregators have listed rate-limited free variants of K2 models (OpenRouter has carried a kimi-k2.6 free listing, for example). Real, but capped hard, sometimes queued, and withdrawn without notice; fine for a curiosity test, wrong for anything that must keep working.
• APIsRouter's free starting credit: issued without a card, spendable on kimi-k2.6 and kimi-k2.7-code at the same catalog rates as paid traffic. The catch is the honest one: it is a starting credit, not an unlimited tier, and after it access is pay-as-you-go.

What Moonshot offers, as precisely as the sources allow.

Moonshot's developer platform at platform.moonshot.ai is pay-as-you-go: the published K2.6 and K2.7-Code rate is $0.95 per million input tokens and $4.00 per million output tokens, per Moonshot's pricing as relayed by launch coverage in June 2026. Around that paid core, third-party pricing guides describe a small new-account trial allowance and a bonus that matches your first recharge at a low threshold; one setup guide also reports that a minimum top-up is required to activate API access at all. We state those as reports rather than facts because Moonshot has revised its onboarding and promotional details repeatedly, and the platform's own documentation is the only current source of truth. The stable, verifiable part is the shape: a metered API with occasional small incentives, not a free tier. The consumer side is different: the Kimi app and web chat run a genuinely free plan with usage limits. If your goal is to try the models conversationally, that costs nothing and needs no key. It just is not an API, and no automation can sit on top of it.

What the free starting credit here actually is.

The route on this site is deliberately simple: a free starting credit, no card required, spendable on any catalog id, including kimi-k2.6 and kimi-k2.7-code at their full 256K context. There is no degraded free endpoint and no separate free-model list; the credit meters the same models at the same rates as paid traffic. When it runs out, access continues as pay-as-you-go top-ups with no subscription, and an idle month costs nothing. To be precise about what this is not: it is not unlimited free Kimi. No one serves a trillion-parameter model without metering it somewhere, and any page promising otherwise is describing a rate limit in optimistic language. What a starting credit buys is a real evaluation: current ids, real rate behavior, and exact per-request token counts in the usage log, before any money moves. The practical pattern is to prototype against the credit, watch what your prompts actually cost, and let those numbers, rather than a pricing page, decide what the workload costs at steady state.

curl https://api.apisrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $APISROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k2.6",
    "messages": [{"role": "user", "content": "Three test cases for a URL parser, as JSON."}]
  }'

The self-hosting route: open weights, one-trillion-parameter hardware.

The K2 family ships open weights under a Modified MIT license; K2.7-Code's checkpoint went up on Hugging Face with the June 12, 2026 release, and K2.6 before it. Legally, this is the most unrestricted free Kimi there is: no rate limits, full data control, and a license that permits serious use. The catch is scale. These are Mixture-of-Experts models reported at around 1 trillion total parameters with roughly 32 billion active per token, and serving one at usable speed is multi-GPU server territory, research clusters and infra teams, not workstations. If that describes your hardware, the weights are the best answer on this page. For everyone else, the checkpoint being free does not make running it free; the electricity alone outruns metered API pricing for any normal workload. The middle path some teams take: a small open model self-hosted for private or bulk traffic, and the hosted K2 ids for the requests that need the real model.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

When free stops being the right answer.

Free routes fail at predictable moments: a mirror's rate cap lands mid-run, a trial allowance expires the week your prototype starts working, a promotion's terms change under you. The signal to move is never ideological; it is the first time a limit costs more time than the metered price would have cost money. Kimi pricing makes that a low bar. At catalog rates a typical chat turn on kimi-k2.6 costs a small fraction of a cent, and workloads that outgrow free routes usually land at single-digit dollars per month. Agent workloads are the exception worth respecting: long tool-calling runs multiply output tokens, which is exactly why evaluating on a metered route with a usage log beats evaluating on a capped mirror that hides what a run would really cost. Because the starting credit and paid traffic share one endpoint and one key, outgrowing free involves no migration: same base URL, same ids, and a balance that tops up instead of a cap that resets.

Which free route fits which situation.

• Trying the models conversationally: the Kimi app's free plan. Zero cost, zero setup, no API.
• A one-off curiosity test from code: a community free mirror, accepted as disposable.
• Evaluating K2.6 or K2.7-Code for real work: the starting credit here, because it runs the exact ids and rate behavior you would ship with.
• Infra teams with multi-GPU clusters and data-locality requirements: the Modified MIT weights.
• Production traffic of any size: metered pay-as-you-go. No free mechanism anywhere is designed to carry it, and the ones that pretend to are the first to throttle.

FAQ