Kimi K2.8 Preview vs Kimi K3: Which One Should You Use?
Kimi K2.8 Preview and Kimi K3 sit on different surfaces: K3 is a public API model (kimi-k3), K2.8 Preview is a Kimi Code route (kimi-for-coding). Compare specs, access limits, pricing and when to use each.
Published 2026-09-17
Kimi K2.8 Preview and Kimi K3 are not two options on the same shelf. K3 is a public API model, and K2.8 Preview is a route inside Kimi Code. On Kimi's open platform the flagship answers to kimi-k3; inside Kimi Code that same flagship answers to k3, while K2.8 Preview keeps the older ID kimi-for-coding.
K2.8 Preview is the lighter, lower-cost tier: every Kimi Code membership tier can use it, including the 1M-token window, and Moonshot describes its overall performance as approaching K3. K3 is the flagship with 2.8 trillion parameters, but it needs Moderato or above, and its 1M window needs Allegretto or above.
Moonshot has published no per-token price for K2.8 Preview, so there is nothing to set beside K3's official $3.00 per million input tokens and $15.00 per million output tokens. On APIMaster you can use kimi-k3 today with one OpenAI-compatible key; K2.8 Preview is not a public API model yet, and we will list it as soon as Moonshot opens it to the open market.
The same model has different IDs on different surfaces
This is the part that trips people up. Kimi runs two separate products, and each one names models its own way.
| Surface | What it is | Kimi K3 ID | Kimi K2.8 Preview ID |
|---|---|---|---|
| Kimi open platform | Pay-as-you-go API for any developer | kimi-k3 |
— not offered |
| Kimi Code | Coding service bundled with a Kimi membership | k3, k3-256k |
kimi-for-coding |
Kimi Code currently exposes four model IDs from three models: k3 and k3-256k (the 1M and 262,144-token variants of K3), kimi-for-coding (K2.8 Preview), and kimi-for-coding-highspeed (K2.7 Code HighSpeed).
Two practical consequences:
- Do not send
kimi-k2.8to an API. That identifier does not exist anywhere. On Kimi's open platform, the only current Kimi models arekimi-k3,kimi-k2.7-code,kimi-k2.7-code-highspeedandkimi-k2.6. - A gateway listing an old Kimi name has not automatically become K2.8. Kimi's open platform discontinued the
kimi-k2series in May 2026 andkimi-k2.5plus themoonshot-v1family in August 2026. None of those IDs now resolve to K2.8 Preview.
What each model is officially positioned for
Kimi K3 is Moonshot's most capable model. Kimi's model list describes it as designed for "frontier intelligence scenarios such as software engineering, knowledge work, and deep reasoning," built on 2.8 trillion parameters with native visual understanding and a 1,048,576-token context window. We cover its API surface, tool calling and operational limits in the Kimi K3 launch guide.
Kimi K2.8 Preview is positioned for everyday development rather than frontier work. Kimi's release notes say it improves coding and agent capability with noticeably better thinking efficiency than K2.7 Code, and is suited to code completion and routine development tasks. It shipped on September 11, 2026 as a full rollout inside Kimi Code — and the model ID did not move, so existing kimi-for-coding configurations picked it up without a client change. Our Kimi K2.8 Preview guide walks through connecting Kimi Code to a third-party key.
The relationship between the two is deliberate: K2.8 Preview is the default working model, K3 is the escalation path.
Spec comparison
Everything below is Kimi's own published specification for Kimi Code, where both models are currently available side by side.
| Kimi K2.8 Preview | Kimi K3 | |
|---|---|---|
| Model ID | kimi-for-coding |
k3 / k3-256k |
| Parameter count | Not published | 2.8 trillion |
| Context window | 1,048,576 tokens | 1,048,576 tokens (k3) / 262,144 (k3-256k) |
| Reasoning effort | low / high / max, default max |
low / high / max, default high |
| Multimodal input | Images and video | Images and video (k3) / images only (k3-256k) |
| Who can call it | Every membership tier | Moderato and above; 1M window needs Allegretto and above |
| Relative consumption | Standard | k3 (1M) costs roughly twice k3-256k |
| Public API price | Not published | $0.30 cached input / $3.00 input / $15.00 output per 1M tokens |
One routing rule is worth memorising because it silently changes which model serves your request: inside Kimi Code, turning thinking off routes both K3 and K2.8 Preview requests to K2.8 Preview without thinking. A request you configured as K3 can therefore be answered by K2.8. This is Kimi Code behaviour; do not assume the same rule applies to the open platform or to a third-party gateway.
What Moonshot has not published
Three gaps are worth stating plainly, because most comparison content fills them in with guesses:
- Parameter count. Kimi published 2.8T for K3. There is no published parameter count for K2.8 Preview.
- Benchmarks. The K2.8 Preview announcement publishes no scores and no reproducible K2.8-versus-K3 evaluation. "Approaches K3" is Moonshot's own assessment, not an independent measurement — ours included.
- Price. See the next section.
Pricing: K3 has a number, K2.8 Preview does not
Kimi's price table lists four models. K2.8 Preview is not one of them.
| Model | Input (cache hit) | Input (cache miss) | Output | Context |
|---|---|---|---|---|
kimi-k3 |
$0.30/M | $3.00/M | $15.00/M | 1,048,576 |
kimi-k2.7-code |
$0.19/M | $0.95/M | $4.00/M | 262,144 |
kimi-k2.7-code-highspeed |
$0.38/M | $1.90/M | $8.00/M | 262,144 |
kimi-k2.6 |
$0.16/M | $0.95/M | $4.00/M | 262,144 |
| Kimi K2.8 Preview | — | — | — | — |
K2.8 Preview is sold as part of a Kimi Code membership, not by the token. Kimi Code shares the same quota as the Kimi membership: the quota refreshes every 7 days, there is an additional rolling 5-hour window, and an extra-usage balance takes over when the quota is exhausted. Kimi describes the extra-usage rate only as "close to the official API price of the Kimi open platform, as shown on the platform" — it does not publish a K2.8 Preview rate.
So the honest comparison is not "K2.8 is cheaper than K3 by X" but "K2.8 is billed as a subscription and K3 as tokens." With that said, three published facts point to K2.8 Preview sitting below K3 on cost:
- Every Kimi Code tier can call K2.8 Preview, while K3 needs Moderato or above.
- K2.8 Preview is positioned for completion and routine development, not frontier work.
- Its closest published sibling on the open platform,
kimi-k2.7-code, is priced at $0.95 input / $4.00 output per million tokens — roughly one third of K3's input price and a quarter of its output price.
The APIMaster route price for K2.8 Preview is a reserved slot. We are not announcing a K2.8 price or confirmed availability. When Moonshot opens the model to the public API and we list the route, this is where the number goes, and you will be able to compare it with K3 card-for-card in the same marketplace.
| Route | Input | Output |
|---|---|---|
| Kimi K3 on APIMaster (today) | $2.25/M — 25% below Kimi's official $3.00/M | $11.25/M — 25% below Kimi's official $15.00/M |
| Kimi K2.8 Preview on APIMaster | Reserved — to be published with the route | Reserved — to be published with the route |
Route prices and channel supply change. Check the live card in the marketplace before you move production traffic.
Both models belong to the same family, so if you are still deciding between Kimi and the other frontier routes, the Kimi API page compares what a single key covers across Kimi, Claude, GPT and DeepSeek.
What this means in practice
Beyond the spec sheet, four operational details decide how the choice feels day to day.
Cache invalidation. Changing model or changing reasoning effort invalidates the context cache, so the next request re-prefills the conversation and costs more. Kimi recommends starting a new session instead of switching back and forth inside a long one. If you plan to compare K2.8 and K3 on the same task, budget for that first expensive request on each.
Permission errors look like authentication errors. A request that exceeds your plan returns 401, not a quota message. Calling k3 below Moderato, requesting the 1M window on Moderato, or calling kimi-for-coding-highspeed below Allegretto all surface as 401. If your key works for one model and fails for another, check the plan before rotating keys.
IDs differ per surface. A k3 configuration is only valid inside Kimi Code. On the open platform the same model is kimi-k3. Copy IDs from the surface you are actually calling, and record provider, endpoint and model ID alongside any benchmark you keep.
Verify, do not assume. A model's own claim about its identity is not proof of which model served a request. This matters more than usual here, because disabling thinking on K3 legitimately hands the request to K2.8 Preview — a routed result that looks like a downgrade if you did not expect it.
How to choose
- Code completion, small refactors and routine development: K2.8 Preview. It is available on every tier, supports the full 1M window, and is the cheaper route.
- Repository-scale changes, long agent runs, deep reasoning and knowledge work: K3. Its 2.8T-parameter flagship tier and 1M context are the reason to pay more.
- You need a public API, SDK compatibility or predictable per-token billing: K3 today, because K2.8 Preview has no public API ID.
- You want the highest throughput inside Kimi Code: neither of the above —
kimi-for-coding-highspeedprovides roughly 6x output speed at 3x consumption, with the same coding capability as K2.7 Code.
What the community is talking about
The two models have very different levels of public discussion, and that is itself a useful signal.
K3 has been heavily discussed since July 2026. On Hacker News, "Kimi K3: Open Frontier Intelligence" reached 2,107 points and 1,214 comments, "Kimi-K3 on HuggingFace" 1,382 points, "Kimi K3 Is Competitive with Fable" 877 points, and "Moonshot AI suspends new subscriptions due to Kimi K3 demand" 284 points — the last one is a reminder that capacity, not just quality, shaped how Moonshot chose to sell its models. Architecture notes, the technical report and local-inference experiments (running K3 in 29 GB of RAM) all attracted substantial threads.
K2.8 Preview has almost no English-language coverage: a Hacker News search for "Kimi K2.8" and "Moonshot K2.8" returns no stories or comments at all, and mainstream third-party gateways do not list the model. Discussion is concentrated in Chinese-language developer channels, which is consistent with a model that currently ships only inside a subscription product. Treat community enthusiasm for K3 as evidence of interest, not as a K2.8-versus-K3 quality verdict.
FAQ
What is the model ID for Kimi K2.8 Preview?
Inside Kimi Code, kimi-for-coding. The ID did not change when the underlying model was upgraded, so existing configurations keep working. Kimi's open platform does not offer K2.8 Preview, so there is no public API model ID for it yet.
Is Kimi K2.8 Preview better than Kimi K3?
Moonshot describes K2.8 Preview as approaching K3 with better thinking efficiency than K2.7 Code, but publishes no benchmark scores for it. K3 remains the flagship 2.8T-parameter model. For a real answer, run your own task — a failing test, a multi-file change, a long-document question — on both and compare correctness, completion time and cost.
Can I call Kimi K3 from an API?
Yes. K3 is a public model on Kimi's open platform as kimi-k3, and it is available on APIMaster with an OpenAI-compatible endpoint. K2.8 Preview is the one that is not callable as a public API model.
Why is the Kimi Code ID k3 when the API ID is kimi-k3?
They are two products with two naming schemes. Kimi Code's four IDs are k3, k3-256k, kimi-for-coding and kimi-for-coding-highspeed; the open platform's current IDs are kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed and kimi-k2.6. Always copy the ID from the surface you are calling.
Does disabling thinking give me Kimi K2.8 Preview?
Inside Kimi Code, yes: requests configured for K3 or K2.8 Preview with thinking turned off are both served by K2.8 Preview without thinking. Do not generalise this rule to the open platform or to a third-party gateway without checking that provider's documentation.
Is Kimi K2.8 Preview available on APIMaster?
Not yet. K2.8 Preview is not a public API model, so there is no route or price to list. APIMaster currently serves kimi-k3, kimi-k2.6 and kimi-k2.7-code. We will list K2.8 Preview as soon as Moonshot opens it to the open market, and the same API key will then cover it.
Sources and verification date
- Kimi Code: model configuration, IDs, limits and routing, checked September 17, 2026.
- Kimi Code: release notes for the September 11, 2026 K2.8 Preview rollout.
- Kimi Code: membership, shared quota, 5-hour window and extra usage.
- Kimi open platform: model list and discontinued models.
- Kimi open platform: model inference pricing.
- Hacker News discussions cited above: 48935342, 49065752, 48999291, 48969291.
- APIMaster live pricing catalog and marketplace cards, checked September 17, 2026. Availability and route prices can change after publication.
Start with Kimi K3 on APIMaster.ai
K2.8 Preview is the model to watch; K3 is the model you can call right now. On APIMaster.ai a single OpenAI-compatible key covers Kimi K3 at 25% below Kimi's official price, alongside Claude, GPT, DeepSeek, Gemini and GLM routes, so you can benchmark the flagship against the rest of your stack before K2.8 Preview lands.
Create an APIMaster account, generate a key in the console, point your client at https://apimaster.ai/v1 and set the model to kimi-k3. Every Kimi channel is fingerprint-checked before it is listed.