What is Kimi K2.7 Code?
Kimi K2.7 Code is Moonshot AI’s coding model, published as one that follows instructions more reliably inside long contexts and completes programming work at a higher success rate, with text, image and video input, a thinking mode, and support for both conversation and agent tasks. The official description places it beside two other identifiers: kimi-k2.7-code-highspeed, the same model on a faster route, and kimi-k2.6, the generation before it. Two relationships are worth fixing before anything else. First, the standard and high-speed identifiers name one model rather than two — the calling guide says so in a single sentence, and the API reference lists exactly those two values for the model field, with the standard one as the default. Second, the previous generation is still on sale and still printed on the same pricing table, which is what makes the price comparison further down this page readable rather than reconstructed. What the standard route does not carry is any published speed figure of its own: the two throughput numbers the vendor gives, about 180 tokens per second and up to 260 on short context, describe the high-speed route only.
What Kimi K2.7 Code is worth using for
Runs you can compare, because sampling cannot drift
Temperature is pinned at 1.0, top_p at 0.95, n at 1, and both penalties at 0.0, and the documentation states that supplying any other value for any of them returns an error. The vendor’s own advice is not to set these fields at all. For a single conversation that is a loss of control. For comparing two runs of the same prompt it removes one variable, because you are not measuring your own sampling choices on top of the model. This workspace exposes none of the five, which happens to land on the same place the official rule does.
I am going to run the same task twice to compare the answers. Before you start, tell me what varies between the two runs on your side and what does not. Then solve the task and list the assumptions you had to supply.
Reasoning that is always on, and always kept
There is no non-thinking mode here. The documentation says plainly that Kimi K2.7 Code does not support one, that the thinking type accepts only the enabled value, and that a disabled value returns an error. Preserved thinking is fixed on for the same reason: the history parameter accepts only all, and passing null or nothing behaves identically. The consequence reaches further than the setting screen, because the reasoning trace from earlier turns stays in context and has to be sent back with each follow-up. That trace is billed as input, so a long session carries its own accumulated reasoning as a running cost.
Keep a running decision log as we work. After each step, restate the constraints you are still honouring, the assumptions you have made since my last message, and anything earlier in this conversation that you are no longer sure about.
A 256K window shared with the previous generation
The window is 256K, printed on the pricing table as 262,144 tokens, and it is identical on the K2.7, the high-speed and the K2.6 rows. So the new generation did not buy reach, and the faster route did not either. What that buys in practice is a session that can hold a repository’s worth of files, a correction round and the reasoning behind both without dropping the earlier part. The workspace above does not display any context figure, because the catalog field it would read is absent, so treat the number as a documented ceiling rather than a gauge you can watch.
I will paste several files and then ask questions across all of them. Read everything before answering, and for each answer tell me which file you relied on and which part of my question the files did not settle.
Screenshots and screen recordings as input
Text, images and video all go in. The calling guide lists nine video formats and recommends capping video at 1080p and images at 4K, on the reasoning that higher resolutions add processing time without adding comprehension. Two constraints are unusual and worth knowing early: image URLs are not supported at all, so the payload must be base64 or an uploaded file, and image and video tokens are counted dynamically, with a separate endpoint for estimating that cost before a request is sent. The one place the official pages disagree with each other is the image format list, which appears as four formats in one document and seven in another; both are given in the guide below rather than merged.
I will attach a screen recording of a bug. Describe what happens in order, separate what you can see from what you are inferring, and then list the two most likely causes with the evidence for each.
Official Kimi K2.7 Code specifications
- Identifier
- kimi-k2.7-code
- Context window
- 262,144 tokens (256K), the same on the K2.6 and high-speed rows
- Default max_tokens
- 32,768
- Sampling parameters
- temperature 1.0, top_p 0.95, n 1, presence and frequency penalty 0.0 — all fixed, other values error
- Thinking
- Always on; only the enabled type is accepted; preserved thinking fixed on; no reasoning_effort
- tool_choice
- auto or none only; any other value errors
- Input modalities
- Text, images and video
Kimi K2.7 Code and Kimi K2.6, the two generations on one table
Every value below is read from a single page — the K2-series table on the Kimi pricing page, which prints a kimi-k2.7-code row, a kimi-k2.7-code-highspeed row and a kimi-k2.6 row one above the other. One official artefact therefore accounts for both columns of every row, which is what makes this readable as a comparison rather than two lists placed side by side. Read the rows in order and the shape of the generation becomes visible: exactly one cell moved. Cache-hit input rose from 1.10 to 1.30 yuan per million tokens. Cache-miss input stayed at 6.50, output stayed at 27.00, the billing unit is unchanged, and the context window is the same 262,144 tokens on every row. The identical rows are kept rather than filtered out, because dropping a row that happens not to differ would be an editorial decision dressed up as a table.
| Official figure | Kimi K2.7 Code | Kimi K2.6 |
|---|---|---|
| Billing unit | 1M tokens | 1M tokens |
| Input price · cache hit | ¥1.30 / 1M tokens | ¥1.10 / 1M tokens |
| Input price · cache miss | ¥6.50 / 1M tokens | ¥6.50 / 1M tokens |
| Output price | ¥27.00 / 1M tokens | ¥27.00 / 1M tokens |
| Context window | 262,144 tokens | 262,144 tokens |
How to use Kimi K2.7 Code online
Bring the whole problem, not a fragment
Thinking is mandatory on this model and its reasoning is preserved across turns, which makes it a poor fit for one-line questions that expect one-line answers and a good fit for a task with a shape. State the outcome you want, the constraints you cannot move, the evidence you want cited, and the condition under which the model should stop and ask rather than guess. Because the earlier reasoning stays in context and is billed like any other input, a long session is also a running cost, so it is worth starting a fresh conversation when the problem changes rather than carrying an unrelated history forward.
Attach what words cannot carry
A screenshot, a diagram or a screen recording is a legitimate input here on the official side, which recommends keeping images under 4K and video under 1080p and reaching for an uploaded file rather than a long data URL when the same asset is referenced more than once. Image URLs are not accepted at all. On this page the attachment control is narrower than the official specification: JPEG, PNG, WebP and PDF, up to 24 MB per file, with no video. A screen recording therefore has to become stills before it can be used here.
Decide between the two routes before a long run
The two identifiers are one model at two prices: the high-speed route costs exactly twice the standard route per token on all three rates, and it changes neither the window nor the capability. So the question is whether a task needs the throughput. Interactive work, agent loops and anything a person is waiting on usually do; work that can run unattended usually does not. There is one more asymmetry worth knowing before choosing, and it is documented rather than obvious: the official client applications do not list the standard route at all, so a subscription allowance is not a way to reach it.
What Kimi K2.7 Code costs, and which official number belongs to which meter
Two official meters apply to this model and they are not interchangeable, so they are kept apart here. The first is per-token billing on the China platform, priced in yuan per million tokens, and it is the table the comparison above is drawn from: cache-hit input 1.30, cache-miss input 6.50 and output 27.00, against 1.10, 6.50 and 27.00 on the K2.6 row that sits directly beneath it, with a context window of 262,144 tokens on both. The second is the same table on the international platform in US dollars, and it repeats the same shape rather than the same numbers: 0.19 for cache-hit input, 0.95 for cache-miss input and 4.00 for output, against 0.16, 0.95 and 4.00 for the previous generation. A third set of figures circulates around this family and belongs to the other route, not this one: the Kimi Code client documentation prints three times quota usage against the high-speed entry, which is a subscription allowance multiplier rather than a per-token rate, and it has no row for the standard variant at all. Three times quota and double the price measure different things, and reading one as the other would be a mistake worth avoiding. One more reading follows from the layout of the pricing table itself: this row contains no conditional pricing, so there is no input-length tier that re-prices a request and no peak-versus-off-peak multiplier. What the page does not contain matters as much. No official page puts an end date on any of these rates, and no official page publishes a promotional or introductory price for this model; both were checked on 7 October 2026. Nothing above is this site’s checkout price. This site bills in points against a catalog reference price of 165.6371 points per million input tokens, 688.0309 per million output tokens and 33.1275 per million cached read tokens, and there is no conversion between points and either currency.
View points and plans ↗Kimi K2.7 Code FAQ
Is Kimi K2.7 Code a different model from Kimi K2.7 Code HighSpeed?
No. The official calling guide states that kimi-k2.7-code-highspeed is the same model as kimi-k2.7-code, served at roughly five to six times the output speed rather than as a separate build. The API reference lists exactly those two values for the model field and describes them as one model with identical parameters, and the pricing table gives both rows the same context window and the same billing unit. What differs is the rate, which is exactly double, and the throughput. The page treats them as one model with two service routes, which is how the vendor writes it.
Can I turn thinking off?
Not on this model, and not through the API either. The documentation states that Kimi K2.7 Code does not support a non-thinking mode, that the thinking type accepts only the enabled value, and that supplying a disabled value returns an error. Preserved thinking is fixed on in the same way: the history parameter accepts only all, and passing null or omitting it behaves identically. There is also no reasoning-effort dial, because that parameter belongs to the K3 model and is listed as unsupported here. The control in the workspace above is therefore present but greyed out.
Did the price go up compared with Kimi K2.6?
One of the three rates did, and only one. On the China platform the cache-hit input price moved from 1.10 to 1.30 yuan per million tokens, roughly eighteen per cent, while cache-miss input stayed at 6.50 and output stayed at 27.00. On the international platform the same cell moved from 0.16 to 0.19 dollars and the other two did not move. The context window is 262,144 tokens on both generations. This is worth stating plainly because the usual summary of a new generation, that it costs more, hides which meter actually changed.
What is the output limit on this page?
Each reply here is requested with at most 4,096 output tokens. The official default for this model is 32,768, and that figure describes the API rather than this chat box. The gap exists because the catalog this site reads carries no maximum-output field for this model, so the composer fallback applies. Treat the larger number as a property of the official endpoint and the smaller one as what this interface actually asks for.
How much context do I get here?
The official window is 262,144 tokens, which the calling guide writes as 256K, and it is identical on the standard, high-speed and K2.6 rows of the pricing table. This interface does not display a context figure at all, because the catalog field it would read is absent, so any number you see quoted for this page is the documented ceiling rather than a gauge that reflects a live session.
Can I upload a video here?
No. The official pages accept video as a first-class input and list nine formats, and the calling guide recommends keeping video under 1080p. This site’s attachment control accepts JPEG, PNG, WebP and PDF only, at up to 24 MB per file. There is no video path to narrow, so the difference is larger than a missing file extension: a screen recording has to be reduced to stills before it can be used here.
Do these rates have an expiration date?
None is published. The China pricing page and the international pricing page were both checked on 7 October 2026, and neither attaches a date, a promotional window or an introductory price to this model’s rates. One line on the pricing page mentions a limited-time arrangement, but it describes file content extraction and file storage rather than model inference, and it names no window at all, so it is not treated here as a price for this model. Read the figures as current at the last edit rather than guaranteed.
Is this the newest Kimi coding model, and can I use it in the official client?
It is not the newest. The platform introduces K3 as its flagship model, 262,144 tokens is a shared window rather than a distinguishing one, and the official client applications list three models across four identifiers: K3, K3-256k, K2.8 Preview and K2.7 Code HighSpeed. The standard route this page describes does not appear in that list under any identifier. So there is no subscription route to it, the three-times-quota figure printed there belongs to the high-speed entry and not to this one, and the API identifier kimi-k2.7-code is the way in.
Sources and model information
Information checked
Start a Kimi K2.7 Code conversation
The chat box above runs this model, and this page is written to be checked rather than trusted. The specifications come from the official calling guide, the comparison comes from the single official pricing table that prints both generations one above the other, and the limits of this workspace come from the code that runs it. Where the two sides disagree — the output ceiling, the context figure, video input, web search — the page says so in the section above instead of quietly writing the official number next to this interface.
Start a chat with Kimi K2.7 Code ↑