Beginner's Guide to Kimi K3: Getting Started in 2026
· 14 min read

Beginner's Guide to Kimi K3: Getting Started in 2026

A practical beginner's walkthrough of Kimi K3: picking the right Kimi product, your first API call, pricing, rate limits, and how it compares to alternatives.

If you are new to Kimi’s models, this is the fastest documented path to a working API call: a $1 top-up on the Kimi API Open Platform, one OpenAI-compatible endpoint, and the model string kimi-k3. This guide is based on Moonshot’s documentation, pricing pages and vendor blog — we did not run the models hands-on. It is for the reader who has used chatbots in a browser, has never held an API key, has Python 3 installed, and wants one successful call tonight on about $5.

What is Kimi K3, in plain terms

Kimi K3 is a 2.8-trillion-parameter model with a 1,048,576-token context window, native understanding of text, images and video, and open weights under the Kimi K3 License — not Apache 2.0. It is Kimi’s previous-generation flagship, not the current one; kimi-k3.1-pro holds that title, per Kimi’s model list.

The plain-language version: Kimi K3 is a large language model — a program you send text to and get text back from — built by Moonshot. “Parameters” are the internal numbers that make up the model; 2.8 trillion of them puts K3 in the 3T class, and Moonshot calls it “the world’s first open-source model in the 3-trillion-parameter class” in the K3 quickstart. Only 104 billion parameters are activated per token, and the vision encoder adds 401 million, per the Kimi K3 model card.

The context window is 1,048,576 tokens. A token is a chunk of text roughly the size of a short word, so a 1M-token window lets you hand the model an enormous document in one request instead of chopping it up. “Native” text, image and video understanding means no separate vision model, and thinking mode is always on: you dial the amount of reasoning rather than turning it off.

How This Guide Was Built

This guide was built from Moonshot’s API documentation, its pricing pages and the vendor blog, cross-checked against the model card on Hugging Face. We verified the per-token prices on Kimi’s pricing page, the tier table on the rate-limits page, the $1 minimum top-up and $5 voucher, the base URL https://api.moonshot.ai/v1, the model strings kimi-k3, kimi-k3.1-pro and kimi-k3.1-pro-highspeed, and the 1,048,576-token context shared by all K3 models.

What we did not test: latency, throughput, uptime or answer quality. No hands-on testing was performed for this article. The benchmark figures below are vendor-run numbers from Moonshot’s technical blog, not independent measurements. For the scored view instead of the walkthrough, read our Kimi K3 review.

Last verified: November 2026

Kimi K3 vs alternatives: which should you pick?

Kimi K3 competes on context length and openness rather than on being the cheapest token — and inside Kimi’s own lineup, kimi-k3.1-pro is the better-value pick at the same endpoint, per Kimi’s pricing page.

ProductPrice (per MTok)Free tierEase for beginnersKey featuresBest fit
kimi-k3$3.00 in / $15.00 out / $0.30 cached inNo free tier in the docs we checked; a $1 minimum top-up makes the model availableEasy — one OpenAI-compatible endpoint1M-token context, image + video understanding, thinking always onBeginners who want the documented open 3T-class model
kimi-k3.1-pro$2.00 in / $10.00 out / $0.20 cached inSameSame setup — change only the model stringCurrent Moonshot flagship, 3T-classPrice-conscious beginners: cheaper per token than kimi-k3
kimi-k3.1-pro-highspeed$4.00 in / $20.00 out / $0.20 cached inSameSame setupSame intelligence at ~2-3x speedReaders who need faster replies
Batch API (kimi-k3 / kimi-k3.1-pro)60% of the standard priceFollows the same paid modelModerate — bulk jobs, not real-time chatSame models at a discountHigh-volume, non-urgent work
DeepSeek deepseek-flash$0.15 in (off-peak, cache miss) / $0.60 outNot in the pages we checkedSeparate provider and account1M context, 384K max output, vision on flash onlyTightest budget that still needs 1M context
DeepSeek deepseek-v4-pro$0.66 in / $1.98 out (off-peak)Not in the pages we checkedSeparate provider and accountDeepSeek’s pro tierDeepSeek users who do not need vision
Claude (Fable / Opus / Sonnet plans)Per-token prices not quoted — the table did not render in our capture; check Anthropic’s pricing pageNot in the pages we checkedSubscription plans plus API creditsMax $100 / Ultra $200 monthly API creditsReaders already inside the Claude ecosystem

If price drives the decision, kimi-k3.1-pro undercuts kimi-k3 at the same endpoint — $2.00 in and $10.00 out per million tokens versus $3.00 and $15.00, per Kimi’s pricing page. If you want the exact model this guide documents, kimi-k3 is the same one-liner, and switching later is a single string change. For a wider field, compare LLMs side by side in our database.

Before you spend anything: the four Kimi products

Four products sit under the Kimi name, and only one sells the pay-as-you-go API access this guide uses — the Kimi API Open Platform, described on Kimi’s product-plans page. Beginners who buy the wrong one end up with a subscription that cannot produce an API key for these models.

  1. Kimi API Open Platform — pay-as-you-go access to the models over an API: top up, get a key, pay per token. This is what an API beginner wants.
  2. Kimi Code — an AI coding assistant on desktop, CLI and IDE plugins, with one subscription covering all three. The important detail: “the Kimi Code API is independent from the API service on this platform, and the two are not interchangeable,” per the same product-plans page. The product site is kimi.ai/code.
  3. Kimi Business — an annual subscription aimed at teams. It is not the route to a personal API key.
  4. Kimi Membership — “four membership subscription tiers”, billed monthly or annually, per Kimi’s product-plans page.

We did not capture prices for Kimi Code, Kimi Business or Kimi Membership, so we are not quoting any — check the vendor pages before you buy. For a walkthrough of the consumer side, see our full Kimi review.

Getting started: your first Kimi K3 call in 7 steps

Seven steps take you from zero to a printed answer: confirm the Open Platform, top up $1, create a key, install the OpenAI client, call the endpoint with kimi-k3, read the split thinking and answer fields, then go multi-turn — per the K3 quickstart.

1. Confirm you are on the right product. You want the Kimi API Open Platform, not Kimi Code. If a page is selling a coding assistant subscription for a desktop app, you are in the wrong place — see the product plans.

2. Top up at least $1. The quickstart is explicit: kimi-k3 becomes available “after a successful top-up (minimum $1)”, per Kimi’s quickstart. Your cumulative top-up also sets your tier and rate limits, so $1 puts you on Tier 0 and $10 moves you to Tier 1.

3. Create an API key and treat it like a password. Generate it in the API key console, paste it into your script once, and never commit it to a public repository.

4. Install the client. One command, copy-paste it:

pip install --upgrade 'openai>=1.0'

If pip is not on your path, the form in Kimi’s documentation works too: python3 -m pip install --upgrade 'openai>=1.0'. You need Python 3.9 or newer, per the quickstart.

5. Make your first call. Save this as first_call.py, replace YOUR_API_KEY, and run it:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url="https://api.moonshot.ai/v1",
)

response = client.chat.completions.create(
    model="kimi-k3",
    messages=[
        {"role": "user", "content": "Explain what a 1M-token context window means in one sentence."}
    ],
)

print(response.choices[0].message.content)

The only two things that make this a Kimi call rather than an OpenAI one are base_url="https://api.moonshot.ai/v1" and model="kimi-k3", both from the K3 quickstart. At $3.00 per million input tokens on Kimi’s pricing page, a test call costs a fraction of a cent.

6. Read the response properly. Thinking is always on. Streaming returns separate reasoning_content deltas and then the final content deltas, per the quickstart: reasoning_content is the model thinking out loud, content is the answer you wanted.

7. Go multi-turn correctly, and tune effort. When you send a follow-up, resend the complete assistant message, not just its content — the quickstart says this for both multi-turn chats and tool calls. Optionally pass reasoning_effort as low, high or max; the default is max, and low is plenty while you are learning. If you later move from scripts to building agents that call tools in a loop, the same rule applies.

Pricing and rate limits, explained for beginners

Kimi K3 is priced per token, and a $1 top-up puts you on Tier 0 with 3 requests per minute — so the two numbers to plan around are the per-million-token price and the request rate, both documented on Kimi’s pricing page and rate-limits page.

Prices, per million tokens (MTok), all with a 1,048,576-token context:

ModelInput / MTokOutput / MTokCached input / MTok
kimi-k3$3.00$15.00$0.30
kimi-k3.1-pro$2.00$10.00$0.20
kimi-k3.1-pro-highspeed$4.00$20.00$0.20

For kimi-k3, cache writes cost $3.00 per MTok on the 5-minute cache and $6.00 per MTok on the 1-hour cache, per Kimi’s pricing page. Output is five times the input price, which is the most useful budgeting fact here: short questions with long answers cost more than long questions with short answers.

Now the limits. These are cumulative-top-up tiers, from the rate-limits page:

TierCumulative top-upConcurrencyRPMTPMTPD
Tier 0$113500,0001,500,000
Tier 1$10151002,000,000unlimited
Tier 2$20401003,000,000unlimited
Tier 3$100502003,000,000unlimited
Tier 5$3,0001003005,000,000unlimited

And a bonus: “when your cumulative recharge reaches $5, you receive a $5 voucher.”

Translated for a first-timer: RPM is the number of separate requests you may send per minute — three at Tier 0, so a loop firing ten calls will fail after the third — and TPM is the tokens you may send and receive per minute, capped at 500,000 on Tier 0. One more trap: limits are enforced per user, not per key, and are shared across all models, per Kimi’s introduction docs.

Two levers cut the bill further. Context caching is automatic — “a default TTL (time-to-live) of 5m” for repeated request prefixes, per the pricing page — so if every request opens with the same long instructions or document, that repeated part bills at $0.30 instead of $3.00 per MTok on kimi-k3; Moonshot reports “a cache hit rate above 90% in coding workloads” on its technical blog. The second lever is the Batch API: kimi-k3 and kimi-k3.1-pro “support batch inference … billed at 60% of the standard price” for work you do not need answered now.

How good is Kimi K3? The honest answer

Moonshot reports Kimi K3 at 67.3 on DeepSWE (the mini-SWE-agent harness on the official leaderboard) and 90.4 on BrowseComp at the full 1M-token context, in its own technical blog — self-reported numbers that we did not measure ourselves.

Architecture in one beginner paragraph: K3 uses a Stable LatentMoE design that “efficiently activates 16 out of 896 experts”, giving roughly 2.5x the scaling efficiency of K2, on Kimi Delta Attention (KDA) plus Attention Residuals, per Kimi’s quickstart. In plain terms: an “expert” is a specialised sub-network, and activating 16 of 896 per token keeps the compute bill down while the model stays huge.

Now the caveat, in the same breath as the numbers. Those results all use reasoning effort max, temperature 1.0, top-p 1.0 and one of three harnesses (Kimi Code, Claude Code, Codex), per Moonshot’s technical blog. They are Moonshot’s measurements on Moonshot’s settings: a signal that the model is strong at long-context and agentic work, not a promise about your task.

One scheduling note: the K2 series — kimi-k2.7-code and kimi-k2.6 — is no longer available to newly registered users, with deprecation expected in late December 2026, per Kimi’s model list. Start on K3-series models and you skip the migration.

Three pro tips

Three habits keep a $5 budget and a Tier 0 account comfortable: set reasoning_effort to low while you are learning, keep a stable long prefix so automatic caching kicks in, and push bulk work through the Batch API at 60% of standard price.

  1. Turn reasoning_effort down while learning. The default is max. For “does my key work” tests and simple questions, low is faster and cheaper, per the K3 quickstart.
  2. Reuse long prefixes. Caching is automatic, so putting your big instruction block or reference document at the start of every prompt — unchanged — gets you the cached-input rate without extra code, per the pricing page.
  3. Queue bulk work through the Batch API. Anything that is not interactive belongs in a batch at 60% of the standard price for kimi-k3 and kimi-k3.1-pro, per the same pricing page.

Common mistakes to avoid

Four mistakes account for most beginner frustration on Kimi: buying Kimi Code when you wanted API access, blaming your script for Tier 0 throttling, ignoring reasoning_content, and trimming the assistant message in multi-turn chats.

  • Buying Kimi Code when you wanted the API. Kimi Code is a subscription for a coding assistant on desktop, CLI and IDE, and “the Kimi Code API is independent from the API service on this platform, and the two are not interchangeable,” per Kimi’s product plans.
  • Assuming your script is broken at three requests per minute. That is Tier 0 throttling, not a bug: 1 concurrency and 3 RPM, per Kimi’s rate-limits page. Add a delay, or top up to $10 for Tier 1.
  • Ignoring reasoning_content. Thinking is always on, and streaming returns separate reasoning_content and content deltas, per the quickstart. Read content for the answer and you will not misread your logs.
  • Trimming the assistant message in multi-turn chats. Resend the complete assistant message, not just its content, or follow-up turns lose context the model expects, per the K3 quickstart.

Kimi K3 FAQ

Short answers to the three questions beginners ask most, based on Kimi’s pricing page and rate-limits page.

Is Kimi K3 free?

No. The docs we checked list no free tier for kimi-k3; the model becomes available after a top-up, and the minimum is $1, which also sets your Tier 0 limits, per Kimi’s quickstart. The weights are downloadable at no charge under the Kimi K3 License, but running them yourself needs hardware, not a $5 budget.

Why does my script stop with rate-limit errors on day one?

Because a $1 top-up puts you on Tier 0: one concurrency and 3 requests per minute, per Kimi’s rate-limits page. A loop that fires ten calls in a second trips the limit immediately. Add a delay between calls, or top up to $10 for Tier 1 and 100 RPM.

Should a beginner start with kimi-k3 or kimi-k3.1-pro?

Either works — same endpoint, same code, one string change. kimi-k3.1-pro is the current flagship and cheaper per token at $2.00 in and $10.00 out, versus $3.00 and $15.00 for kimi-k3 on Kimi’s pricing page. Pick kimi-k3.1-pro if price matters; pick kimi-k3 if you are following this quickstart exactly.

Where to go next

Start with our full Kimi K3 review for the scored breakdown, then compare LLMs side by side if you are still choosing a provider.

Three sibling guides pair well with this one: run models locally with Ollama once the API bill feels real, what fits in 8GB of VRAM if you are sizing hardware, and Gemini plans and usage limits for a second provider’s pricing model to compare against. Get the one successful call done first — everything else is tuning.

  • ToolBrain — tool reviews, LLM comparisons, and AI workflow guides
  • CodeIntel Log — code quality, debugging, and software engineering benchmarks

Cross-links automatically generated from None.

Back to all posts