Fact-checked: August 5, 2026 · Evidence level: official documentation and official model cards, plus one separately reported single-pass K3 API pilot; no independent cross-model comparison or reliability test.
Kimi’s current model line serves three different needs. Kimi K3 is the flagship for the most demanding coding, reasoning and knowledge-work tasks. Kimi K2.7 Code is the lower-cost, coding-focused choice. Kimi K2.6 remains an active general-purpose model and is the only one of these three that lets API users turn thinking off.
This page compares the models rather than teaching API implementation. For code examples and authentication, use our Kimi API guide. For billing terms and cost examples, see Kimi API pricing.
Quick model chooser
| Your priority | Recommended model | Why | Main trade-off |
|---|---|---|---|
| Highest overall capability and up to 1M context | Kimi K3 | Kimi’s current flagship, built for long-horizon coding, deep reasoning and end-to-end knowledge work | Much higher API output price; thinking cannot be disabled |
| Repository-scale coding at a lower token price | Kimi K2.7 Code | Dedicated coding model with a 256K context window and mandatory thinking | Less general-purpose than K3 or K2.6; no non-thinking mode |
| General writing, analysis, vision or faster non-thinking responses | Kimi K2.6 | General-purpose multimodal model with optional thinking | No longer the flagship and materially behind newer models in Moonshot’s reported evaluations |
| Lower latency from the K2.7 family | K2.7 Code HighSpeed | Same underlying K2.7 Code model according to Kimi, with faster output | Double the standard K2.7 API token prices and separate Kimi Code plan restrictions |
Short answer: start with K2.6 for mixed, cost-sensitive workloads; choose K2.7 Code for sustained programming work; move to K3 when the task benefits from the strongest model or more than 256K tokens of context.
Kimi model comparison
The table below covers the Global Kimi API Platform. Kimi Code uses different model IDs and a membership quota rather than the same pay-as-you-go balance.
| Model | Current API status | Official API model ID | Context window | Hosted API inputs | Thinking | Cache-hit input / 1M tokens | Cache-miss input / 1M tokens | Output / 1M tokens |
|---|---|---|---|---|---|---|---|---|
| Kimi K3 | Active; current flagship | kimi-k3 | 1,048,576 tokens | Text, image and video | Always on; low, high or max effort | $0.30 | $3.00 | $15.00 |
| Kimi K2.7 Code | Active; coding specialist | kimi-k2.7-code | 262,144 tokens | Text, image and video | Always on | $0.19 | $0.95 | $4.00 |
| Kimi K2.7 Code HighSpeed | Active | kimi-k2.7-code-highspeed | 262,144 tokens | Text, image and video | Always on | $0.38 | $1.90 | $8.00 |
| Kimi K2.6 | Active; general-purpose | kimi-k2.6 | 262,144 tokens | Text, image and video | On by default; can be disabled | $0.16 | $0.95 | $4.00 |
Prices are official list prices checked on August 4, 2026, exclude applicable taxes and may change. “Cache hit” is not a setting you can assume: Kimi’s automatic context cache must actually match eligible reused context. Check the live Kimi model list and official pricing pages before budgeting production traffic.
Do not mix up API IDs and Kimi Code IDs
The same model can have a different identifier depending on the product.
| Model version | Kimi API ID | Kimi Code membership ID |
|---|---|---|
| Kimi K3, up to 1M | kimi-k3 | k3 |
| Kimi K3, fixed 256K in Kimi Code | Not listed as a separate Global API model | k3-256k |
| Kimi K2.7 Code | kimi-k2.7-code | kimi-for-coding |
| Kimi K2.7 Code HighSpeed | kimi-k2.7-code-highspeed | kimi-for-coding-highspeed |
| Kimi K2.6 | kimi-k2.6 | Used as a fallback when thinking is disabled; not one of the four current selectable Kimi Code IDs documented on August 4, 2026 |
Entering a display name such as “Kimi K3” where a model ID is required can cause a request to fail. Kimi Code also has its own base URL, API keys, plan entitlements and quota rules. A Kimi membership is not the same product as pay-as-you-go access on the Kimi API Platform.
Which Kimi model is best for each workload?
Choose Kimi K3 for the hardest or longest tasks
K3 is the clearest choice when model capability matters more than token price. Moonshot positions it for large-codebase engineering, deep reasoning and multi-step knowledge work. Its 1,048,576-token window is four times the capacity of K2.7 Code or K2.6.
Capacity is not a promise that every fact buried in a million-token prompt will be recalled accurately. Long prompts can increase latency and cost, and agent frameworks may compact history before the model reaches its advertised limit. Use the smallest context that preserves the evidence required for the task.
Read the full Kimi K3 specifications, pricing and limitations.
Choose Kimi K2.7 Code for sustained software engineering
K2.7 Code is purpose-built for coding agents and long-horizon software tasks. Moonshot reports that it follows instructions more reliably than K2.6 in long contexts and uses about 30% fewer thinking tokens on average. Those are vendor-reported findings, not results reproduced by Kimi AI Guide.
K2.7 Code always thinks. That makes it a poor fit when you specifically want a short, non-reasoning response, but it is a natural candidate for repository exploration, debugging, refactoring and multi-step tool use. The standard API version costs the same as K2.6 for cache-miss input and output tokens.
Read our Kimi K2.7 Code model guide.
Choose Kimi K2.6 for flexible general-purpose work
K2.6 is still listed as active by the official Kimi API documentation. It supports the same 256K context window and hosted text, image and video inputs as K2.7 Code, but it is designed as a more balanced general-purpose model. It also supports both thinking and non-thinking modes.
That optional thinking mode is useful when one application mixes hard analysis with routine extraction, rewriting or classification. It can avoid unnecessary reasoning on simpler requests, although response speed and quality still need to be measured in your own environment.
Read our Kimi K2.6 model guide.
A practical API cost comparison
Kimi bills input and output tokens separately. A simple estimate is:
estimated cost = (cache-hit input × hit rate + cache-miss input × miss rate + output × output rate) ÷ 1,000,000
For a request with 1,000,000 cache-miss input tokens and 100,000 output tokens, the list-price estimate is:
| Model | Input cost | Output cost | Estimated total |
|---|---|---|---|
| Kimi K3 | $3.00 | $1.50 | $4.50 |
| Kimi K2.7 Code | $0.95 | $0.40 | $1.35 |
| Kimi K2.6 | $0.95 | $0.40 | $1.35 |
This example excludes taxes, paid tools, retries and any membership charges. It also assumes the entire input is billed as a cache miss. Actual requests can contain a mix of cached and uncached input tokens.
What “open” means here
Moonshot has released model weights for all three model versions:
- K3:
moonshotai/Kimi-K3, under the Kimi K3 License. - K2.7 Code:
moonshotai/Kimi-K2.7-Code, under a Modified MIT License. - K2.6:
moonshotai/Kimi-K2.6, under a Modified MIT License.
Open weights do not make local deployment lightweight. K3 has 2.8 trillion total parameters and 104 billion activated parameters; K2.7 Code and K2.6 each have 1 trillion total parameters and 32 billion activated parameters. Hardware, quantization, inference software and the serving template can all change output quality. Results from the official hosted API should not automatically be attributed to an unrelated third-party deployment.
Are older Kimi K2 model IDs deprecated?
Yes—but the naming is easy to misread. Kimi’s current model list says the older IDs kimi-k2-0905-preview, kimi-k2-0711-preview, kimi-k2-turbo-preview, kimi-k2-thinking and kimi-k2-thinking-turbo were discontinued on May 25, 2026.
That deprecation notice does not list kimi-k2.6. K2.6 appears in the active multimodal-model section alongside K3 and K2.7 Code as of this page’s verification date. Applications should still request the live model list rather than hard-code assumptions indefinitely.
Our verification and testing status
We separate three evidence levels:
- Verified documentation facts: model IDs, published context limits, supported request modes, listed prices and current documentation status were checked against official Kimi sources on August 4, 2026.
- Vendor-reported evaluation: benchmark scores and relative performance claims come from Moonshot’s model cards or technical posts. They are labelled as vendor-reported because harnesses, tools and reasoning settings affect results.
- Independent testing: one single-pass K3 API pilot completed August 5, 2026 under frozen protocol v1.1. Its 15/15 completed calls included 8/8 exact tool cases, a code repair that passed 5/5 fixed tests, 6/6 vision fields and 20/20 context fields across five requests from approximately 32K to 950K prompt tokens. The usage-derived cost was $5.687397; a later balance check showed a $5.68739 deduction. This does not compare K3 with K2.7 Code or K2.6 and is not a reliability estimate. Read the complete Kimi K3 pilot report for raw evidence and limits.
Our planned comparison will use the official API, fixed prompts, disclosed parameters, repeated trials and recorded token usage. See our testing methodology for the rules that govern future results.
Limitations that apply to every model
- A large context window does not guarantee perfect retrieval or reasoning across the entire window.
- Tool-calling success depends on the model, tool schemas, agent loop and error handling—not the model alone.
- Image and video inputs are converted into billable tokens; token use varies with resolution and video keyframes.
- API models do not browse by default. Kimi’s current documentation provides official web-search tools, but tool availability, supported model/mode combinations, token use and the unresolved per-call price must be verified separately from the base model.
- Model availability, prices and rate limits can change independently of this page.
- Outputs can be incorrect. Verify code, citations, financial figures, medical statements and other consequential claims with appropriate sources and tests.
Frequently asked questions
What is the latest Kimi model?
Kimi K3 is the current flagship in Kimi’s official model list as of August 4, 2026. “Latest” does not automatically mean best value: K2.7 Code can be a better fit for cost-sensitive coding, while K2.6 offers optional non-thinking responses.
Which Kimi model has a 1M context window?
Kimi K3 supports a 1,048,576-token context window through the Global Kimi API. Kimi Code plan entitlements can limit access to the full window, so check the product-specific rules rather than assuming membership access matches API access.
Is Kimi K2.6 deprecated?
No. The official model list still places kimi-k2.6 in its active multimodal-model section. Several older K2 preview and thinking IDs are deprecated, which may be the source of the confusion.
Is Kimi K2.7 Code only for code?
It can accept text, images and video and can reason about general prompts, but Moonshot explicitly positions it as a coding-focused model. K2.6 is the vendor’s more general-purpose choice; K3 is the higher-capability flagship.
Can I turn thinking off?
You can disable thinking for K2.6 through the API. K2.7 Code and K3 require thinking; K3 instead lets API users select low, high or max reasoning effort.