Kimi K3
API pricing, subscription options, coding setup, benchmark context, free access, and the limitations you should know before choosing Moonshot AI's flagship model.
Kimi K3 is a 2.8T-parameter, open-weight multimodal model with a 1M-token context window, built for long-horizon coding, knowledge work, and agentic reasoning.
What is Kimi K3?
Kimi K3 is Moonshot AI's most capable model as of August 2026. It combines a sparse Mixture-of-Experts architecture, native visual understanding, persistent reasoning, and a context window large enough to hold extensive repositories or document collections.
The model has 2.8 trillion total parameters, with approximately 104 billion parameters activated per token. Its architecture uses Kimi Delta Attention, Gated MLA, Attention Residuals, and 896 experts, selecting 16 experts for each token. Moonshot distributes the weights under its own Kimi K3 License, so “open-weight” should not be read as automatically equivalent to an OSI open-source software license.
- Developer
- Moonshot AI
- Model type
- Open-weight multimodal MoE
- Total / active parameters
- 2.8T / 104B
- Context window
- 1,048,576 tokens
- Official API model ID
kimi-k3- Inputs
- Text, images, and video
The API exposes low, high, and max through the top-level reasoning_effort field. max is the documented default.
Kimi K3 price and API cost
The official global API uses pay-as-you-go token billing. Input that hits the automatic context cache costs one tenth of a cache miss. Output is priced separately and includes the model's generated response.
| Model | Cached input | Uncached input | Output | Context |
|---|---|---|---|---|
kimi-k3 | $0.30 / 1M tokens | $3.00 / 1M tokens | $15.00 / 1M tokens | 1,048,576 tokens |
Prices are from the global Kimi API Platform on August 5, 2026 and exclude applicable taxes. Regional platforms, exchange rates, promotions, and third-party providers may differ.
What does one request cost?
One million uncached input tokens plus 100,000 output tokens would cost about $4.50: $3.00 for input and $1.50 for output. If all one million input tokens hit the cache, the same token volumes would cost about $1.80. Real bills depend on actual tokens and cache behavior.
Kimi K3 subscription plans
Kimi's hosted workspace has four paid membership tiers. Monthly prices are $19, $39, $99, and $199; annual billing lowers their effective monthly prices to $15, $31, $79, and $159. These plans bundle different levels of Agent and Kimi Code credits. They are separate from API pay-as-you-go billing.
The downloadable weights are available under the Kimi K3 License, but hosting a 2.8T model is not practical on ordinary consumer hardware. The official API is paid. Kimi has a free workspace tier, while Kimi Code documentation requires at least Moderato for K3 access and Allegretto for its full 1M context.
How to use the Kimi K3 API
Kimi's API is compatible with the OpenAI API format. Create a key in the Kimi API Platform, keep it in an environment variable, use the global base URL, and select kimi-k3.
- 01
Create an API key
Sign in to the Kimi API Platform and create a key from the API Keys page.
- 02
Set the environment variable
Store the secret as
MOONSHOT_API_KEY; never commit it. - 03
Call the chat completions endpoint
Choose the model and explicitly set the reasoning effort for your workload.
import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.MOONSHOT_API_KEY,
baseURL: "https://api.moonshot.ai/v1",
});
const response = await client.chat.completions.create({
model: "kimi-k3",
reasoning_effort: "high",
messages: [
{ role: "user", content: "Review this function for edge cases." },
],
});
console.log(response.choices[0].message.content);For multi-turn conversations and tool calls, Kimi says to pass the previous assistant message back as returned, including reasoning_content andtool_calls. Dropping the thinking history can destabilize later responses.
Use Kimi K3 with OpenCode
OpenCode has built-in Moonshot AI authentication. The official Kimi guide was tested against OpenCode 1.18.3, so menu labels may change in later releases.
opencode auth login
# Select "Moonshot AI" and paste your Kimi API key
opencode
# Run /models and select Kimi K3
# Run /variants to choose low, high, or maxK3 defaults to maximum reasoning effort. For routine code completion or smaller edits, a lower effort setting can reduce latency and output use. The API account needs a positive balance; Kimi notes that new-user verification vouchers cannot be used for K3.
Kimi K3 or Kimi K2.7 Code?
Choose K3 for large-repository analysis, difficult multi-file work, long autonomous runs, or tasks mixing vision and code. Kimi K2.7 Code has a smaller 256K context and materially lower API pricing, making it the more economical default for ordinary coding tasks where K3's extra capacity is unnecessary.
Kimi K3 benchmark results
Moonshot's published evaluation reports strong performance across coding, agentic, reasoning, and visual tasks. Selected K3 results are reproduced below to make the numbers easy to inspect, not to claim that unlike harnesses are directly interchangeable.
| Benchmark | Kimi K3 score | Area | Important context |
|---|---|---|---|
| Terminal-Bench 2.1 | 88.3 | Coding agent | Kimi Code harness |
| FrontierSWE | 81.2 | Software engineering | Dominance score |
| ProgramBench | 77.8 | Programming | Vendor-reported evaluation |
| BrowseComp | 91.2 | Agentic browsing | Context compaction at 300K |
| SpreadsheetBench 2 | 34.8 | Spreadsheet tasks | Claude Code harness |
| GPQA Diamond | 93.5 | Reasoning | Max effort, temperature 1.0 |
These are Moonshot AI's published scores, not independent tests run by SheetsX. The official report uses different agent harnesses for some models and includes benchmark-specific caveats. Compare the full footnotes before drawing conclusions.
Is Kimi K3 worth using?
Kimi K3 is most compelling when a task genuinely benefits from very long context, persistent reasoning, multimodal input, and sustained tool use. It is harder to justify for short, price-sensitive requests.
- Best forLarge codebases, research corpora, autonomous coding, and complex agents.
- Strengths1M context, open weights, native vision, tool use, and strong coding scores.
- Trade-offsHigh output price, always-on reasoning, large deployment requirements.
- Not ideal forSimple chat, low-latency completion, or commodity extraction workloads.
Known limitations
Moonshot flags three practical issues. First, K3 can become unstable when a client fails to preserve its earlier thinking history or switches models mid-session. Second, it can act too proactively on ambiguous instructions, so agent permissions and system constraints should be explicit. Third, Moonshot itself says the overall user experience still trails the strongest proprietary models it compared against.
We reviewed Moonshot's technical report, product documentation, pricing, and benchmark notes. We did not run an independent Kimi K3 benchmark suite, so this page does not present vendor scores as our own measurements.
How can you use Kimi K3?
| Route | Billing | Use it when |
|---|---|---|
| Kimi workspace | Free tier and paid membership | You want chat, research, agents, and Kimi's own interface. |
| Kimi API | Pay per token | You are building an application or production agent. |
| Kimi Code | Membership quota | You want a terminal or IDE coding agent. |
| Open weights | Infrastructure cost | You have large-scale inference infrastructure and need deployment control. |
Need model-powered workflows in Google Sheets?
SheetsX brings AI generation, analysis, translation, and structured spreadsheet workflows into the sheet you already use.
Explore SheetsX featuresKimi K3 FAQ
What is Kimi K3?
Kimi K3 is Moonshot AI's open-weight flagship model for long-horizon coding, agentic knowledge work, reasoning, and multimodal tasks. It has 2.8 trillion total parameters and a 1,048,576-token context window.
How much does the Kimi K3 API cost?
The official global API price is $0.30 per million cached input tokens, $3.00 per million uncached input tokens, and $15.00 per million output tokens, excluding applicable taxes.
Is Kimi K3 free?
The model weights are available under the Kimi K3 License, but the hosted Kimi K3 API is pay as you go. Kimi has a free membership tier, while Kimi Code access to K3 requires a paid Moderato plan or above according to its documentation.
Does Kimi K3 have an API?
Yes. Use the model ID kimi-k3 through the Kimi API Platform. The service is compatible with the OpenAI API format and also provides Anthropic-compatible access for supported workflows.
Can I use Kimi K3 with OpenCode?
Yes. OpenCode includes Moonshot AI as a provider. Add your Kimi API key with opencode auth login, start OpenCode, then choose Kimi K3 from the models list.
Is Kimi K3 good for coding?
Kimi K3 is designed for long-running software engineering and tool use. Moonshot AI reports strong results on Terminal-Bench 2.1, FrontierSWE, ProgramBench, and other coding benchmarks, but those scores should be read with their harness and test-setting notes.
What is the difference between Kimi K3 and Kimi K3 256K?
In Kimi Code, k3 uses up to a 1M-token context while k3-256k is capped at 256K. Kimi says the 256K option produces the same results within that context range and consumes about half as much subscription quota.
Who created Kimi K3?
Kimi K3 was created by Moonshot AI, the company behind the Kimi product family and Kimi API Platform.
Sources and editorial notes
Technical claims, prices, availability, and limitations were checked against Moonshot AI and Kimi's official materials on August 5, 2026. Pricing and product access can change; verify current terms before purchasing or deploying.
- Kimi K3: Open Frontier IntelligenceKimi
- MoonshotAI/Kimi-K3Moonshot AI on GitHub
- Kimi K3 API quickstartKimi API Platform
- Kimi K3 API pricingKimi API Platform
- Use Kimi models in OpenCodeKimi API Platform
- Kimi membership pricingKimi Help Center
- Kimi Code model configurationKimi Code Docs