Kimi / Model guide
Kimi models: which version should you use?
Kimi’s model names mean different things in its chat app, coding tools and API. Use this guide to choose a suitable model, check where it is available and avoid paying for the wrong kind of access.
By China AI Chat Team · Sources checked · Independent guide
Quick answer: consider K2.6 for everyday app questions or an API that needs optional thinking; K3 for larger-context work and complex app workflows; and K2.7 Code for coding through the public API. K2.8 Preview is currently a Kimi Code option. These are starting points based on documented capabilities, not a tested ranking of the models.
Which Kimi model should you choose?
Choose your route first: the Kimi app, Kimi Code, the public API or downloaded weights. Check the access and allowance for the route you need; availability in one does not automatically grant access through another.
| Your goal | Start by evaluating | Decision that matters |
|---|---|---|
| Ask questions or draft text in the app | K2.6 | Does its answer meet your checks? Escalate difficult tasks rather than assuming every question needs the highest effort. |
| Create an editable presentation, document or spreadsheet | K3 in the relevant app workflow | Verify the exported file and its contents, not just the chat reply. |
| Coordinate a large batch of app tasks | K3 Cluster, if available | Check task scope, account access and credit use before expanding the batch. |
| Change code in a terminal or editor | The K2.8 Preview or K3 options available in Kimi Code | Review the patch and run your tests; confirm the actual Code selection. |
| Build a coding feature using the public API | kimi-k2.7-code, compared with your existing baseline | Measure completed tasks, tool handling and total cost. Consider HighSpeed only if model generation is the bottleneck. |
| Send a request exceeding the 256K API options | kimi-k3, or a smaller focused input | Include room for completion in the context budget. A larger window does not guarantee better retrieval. |
| Use the public API without a thinking phase | kimi-k2.6 | It exposes a non-thinking mode; the current K3 and K2.7 Code API models do not. |
Selection advice is our interpretation of the official specifications below. We have not run a controlled comparison of all these models. For membership and API costs, use the separate Kimi pricing guide.
Kimi app: K2.6, K3 and K3 Cluster
The official app selection guide lists K2.6 for quick text conversations, K3 for complex work and editable files, and K3 Cluster for work distributed across parallel agents. Cluster is an app option; do not turn its display name into an API model ID.
The guide lists Standard and Advanced thinking intensity for K2.6, with Extreme also available for K3 and K3 Cluster. Higher effort can use more tokens. Start with the level your task needs, and check the result before increasing effort.
App limits are separate: the help guide describes about 128K tokens for a K2.6 chat turn, rather than its 256K API capacity. For K3, the current international membership table places the up-to-1M conversation benefit on Max and Ultra. Check your actual model picker and membership screen; Code and API access differ.
Use our Kimi overview for official access. To judge the file workflow, see our Slides export checks, Docs layout correction and Sheets recalculation checks. Those are dated app trials, not evidence that one underlying model beats another.
Current Kimi API models compared
The international Open Platform catalog lists these four active IDs. All four support text and visual input; image or video understanding is different from image or video generation. The limits below describe the hosted API, not the consumer app.
| Exact API model ID | Context capacity | Thinking control | Main selection factor |
|---|---|---|---|
kimi-k3 | 1,048,576 tokens | Always on; low, high or max effort | Larger context; evaluate its higher token cost against the value of completing the task. |
kimi-k2.7-code | 262,144 tokens | Always on | Coding-focused candidate for an application or agent. |
kimi-k2.7-code-highspeed | 262,144 tokens | Always on | Faster-serving K2.7 Code option with separate rates; measure end-to-end time. |
kimi-k2.6 | 262,144 tokens | On by default; can be disabled | General-purpose option when mode control or lower token rates matter. |
Sources: K3 documentation, K2.7 Code documentation, K2.6 documentation and the API specifications and rates. These are provider specifications, not our latency or accuracy measurements.
K3 versus K2.6: the practical API differences are context capacity, whether thinking can be switched off, and token rates. Both accept visual input. K3’s larger capacity alone does not justify moving every short task to it. Compare the cost of a completed, correct result, including reasoning, repeated attempts and any tools.
Where is K2.8? K2.8 Preview is documented in Kimi Code below, but the checked public API catalog does not list a standalone K2.8 ID. A higher version number in one product does not establish availability in another. Use the API setup guide to match the correct account, key and endpoint.
Kimi Code model names and K2.8 Preview
Code has its own model IDs. According to the official Code configuration reference, the current mappings are:
| Code ID | Current model | Documented context / input note |
|---|---|---|
k3 | K3 | Up to 1M for eligible memberships; image and video input. |
k3-256k | K3, 256K variant | 256K; image input, no video. |
kimi-for-coding | K2.8 Preview | Up to 1M; image and video input. |
kimi-for-coding-highspeed | K2.7 Code HighSpeed | 256K; image and video input. |
Turning thinking off can change the served model. Code documents that disabling thinking for the K3 series or K2.8 Preview routes the request to non-thinking K2.8 Preview. That does not mean public API K3 supports non-thinking requests.
The Code reference separates new and legacy plans. K3 access begins at Plus, or legacy Moderato, with 256K context at that level; its 1M context begins at Pro, or legacy Allegretto. For the other Code options, K2.8 Preview is listed from Plus (legacy Andante), and HighSpeed from Pro (legacy Allegretto). The current purchase page uses Plus, Pro, Max and Ultra, while some help tables retain musical names. Check the Code reference and your account’s available selections rather than assuming that an old plan has the same benefits as a similarly priced new one.
Use /model in the official CLI to inspect selections. Names such as kimi-for-coding are aliases whose underlying model can change. Our Kimi Code guide and recorded repair cover setup and one bounded task; they are not a benchmark of today’s complete model menu.
K3 context and access: app, Code or API?
Context is the space available for the request and completion. Instructions, chat history, tool definitions and media take part of that space. A 1M context specification is not a promise of a 1M-token answer, a fixed number of pages, or perfect recall.
| Where you use K3 | Long-context capacity | Access to check |
|---|---|---|
| Kimi app conversation | Up to 1M tokens | Max or Ultra in the current international membership table. |
Kimi Code: k3 | Up to 1M tokens | New Pro and above, or legacy Allegretto and above. Plus / legacy Moderato provides 256K; k3-256k stays at 256K. |
Public API: kimi-k3 | 1,048,576 tokens | A separate Open Platform account and API billing. An app subscription is not an API token balance. |
Sources: the current app membership table, Code model reference and API catalog, checked on 1 October 2026. Confirm the benefits shown for your account before choosing a plan; region or account offers can differ.
For API K3, the documented default reasoning effort is max. An explicit output allowance must accommodate thinking as well as the visible answer. A response that reaches its limit before answering is an incomplete result. Code’s K3 default is high, and its routing also differs. Record the product as well as the model name when comparing runs.
Before choosing a larger window, decide whether the whole document is necessary. For a question about one policy, supplying the relevant passage and its surrounding context may be easier to verify than repeatedly sending an entire archive. For document workflows, our Kimi PDF guide explains source checking.
Can you download Kimi model weights?
Yes, official releases include downloadable checkpoints. Start with Moonshot AI’s model cards for Kimi K3, Kimi K2.7 Code and Kimi K2.6. Read the exact checkpoint’s deployment instructions, supported runtimes and license before choosing this route.
Do not assume that downloadable weights mean an unrestricted license or an easy laptop installation. K3 uses the Kimi K3 License. Hosting a checkpoint yourself also means arranging the serving system and tools; it does not reproduce all the Kimi app’s file-creation workflows. We have not deployed these checkpoints for this guide.
What happened to K2.5, K2 and Moonshot v1?
The official API catalog marks K2.5 and the Moonshot v1 series discontinued on 31 August 2026, and the older K2 API series on 25 May 2026. The kimi-latest alias is also retired. These are notices about that provider’s API service, not declarations that downloaded weights or every third-party endpoint stopped working.
A model can still appear on a release or marketing page after its hosted API ID is retired. Use the active API catalog and retirement notices to choose an integration target. When migrating an old tutorial, check the exact model ID, supported thinking settings, output-limit field and tool-history requirements together.
A practical check before committing to a model
Use this worksheet to narrow the options before testing answer quality. We have not performed it on every Kimi model.
| Check | Evidence to keep | When to reject or change the setup |
|---|---|---|
| 1. Access and identity | Product, provider, exact selection or ID, date and available entitlement. | The required option is unavailable, retired or belongs to a different API. |
| 2. Required input and output | A representative input and the exact result format you need. | The chosen route cannot accept the media or deliver the required file. |
| 3. Completion quality | Expected facts, source passages, test results or file checks defined before the run. | The answer is truncated, unsupported, or produces a file that fails those checks. |
| 4. Total effort and cost | Completed-task time, token or credit usage, retries and manual corrections. | The result needs more correction or expense than your task permits. |
For a document question, require the answer and a verifiable supporting passage. For a code change, require a reviewed patch, passing relevant tests and no unrelated edits. Repeat representative cases before choosing a default; one successful response cannot establish a reliable success rate.
Our existing Deep Research trial and file guides can help you design checks. Their account conditions and dates remain part of the evidence. For comparisons with other assistants, use the comparison directory; those app tests should not be relabelled as a controlled K3-versus-K2.6 evaluation.
Continue with Kimi
Kimi overview and official access · API setup and troubleshooting · Membership and API pricing
ChinaAI is independent of Kimi and Moonshot AI. This page combines dated official specifications with our selection guidance; it adds no new inference benchmark. Read our research methodology or report a correction.