Use K2.8 Preview if you already pay for a Kimi plan with Kimi Code and want long-context coding help that costs less quota than K3. It reads image and video input, but no benchmark has been published for it, so judge it on your own test tasks.
What changed
Kimi K2.8 Preview is Moonshot AI's coding model, rolled out in Kimi Code in September 2026 at the kimi-for-coding model ID with a million-token context window. Moonshot describes its performance as close to Kimi K3 but has published no benchmark score for it.
Provider: Moonshot AI · Family: Kimi
More about Moonshot AI on HokAI
Context window: 1,048,576 tokens
Input modalities: text, image, video · Output: text, tool-calls, code
About Kimi K2.8 Preview
Kimi K2.8 Preview is a coding and agentic model from Moonshot AI. Its release date is September 11, 2026, when Moonshot's Kimi Code "What's New" page announced it as fully rolled out at the existing kimi-for-coding model ID, the route that previously served Kimi K2.7 Code. Clients and third-party tools needed no configuration change to receive it. In Moonshot's lineup it sits below the flagship Kimi K3, released July 16, 2026 with 2.8 trillion parameters.
Benchmarks: Moonshot has published no benchmark score for K2.8 Preview as of October 10, 2026, and no Artificial Analysis page exists for it. The vendor's only performance statement is qualitative: performance close to K3, with coding and agent capabilities improved across the board and significantly more efficient thinking than K2.7 Code. For scale, Moonshot's July 16, 2026 K3 tech blog reported 93.5% on GPQA Diamond for K3, but that figure belongs to K3 and does not carry over to this checkpoint. Treat any K2.8 score you see quoted elsewhere as unverified until Moonshot or an independent evaluator publishes one with a method.
What the Kimi Code model configuration page does confirm: a 1,048,576-token context window, image and video input, and the same low, high and max reasoning-effort levels K3 uses inside Kimi Code, with max as the K2.8 default where K3 defaults to high. Requests sent with thinking turned off, to either K3 or K2.8 Preview, are answered by K2.8 Preview in no-thinking mode, so this model also serves as the fast path behind the flagship.
Moonshot has not published a model card, parameter count or architecture description for K2.8 Preview. Because it replaced K2.7 Code at the same model ID, it most likely shares that model's Mixture-of-Experts transformer design (1 trillion total parameters, 32 billion active per token), but that is an inference from the shared endpoint, not a confirmed specification.
Access runs through a Kimi membership, not per-token billing. On the new plans, kimi-for-coding is available from Plus upward; on legacy plans, from Andante upward. Moonshot's Open Platform pricing page and platform changelog list K3, K2.7 Code and K2.6 but no K2.8 line as of October 10, 2026. The Kimi Code API does accept an API key from the Kimi Code Console over OpenAI-compatible and Anthropic-compatible base URLs, so the model can run inside third-party agents such as Claude Code against the same membership quota.
Who should not pick it: teams that need a documented, independently benchmarked model before adoption, a pinned model version, or open weights. Kimi K2.7 Code shipped open weights under a Modified MIT license, and K3 carries a published benchmark table; for direct pay-as-you-go API billing, DeepSeek V4 and GLM-5.2 are listed with per-token prices. For a broader shortlist, see the best AI coding assistants and the Kimi comparison with ChatGPT.
Pricing
Not billed per token. K2.8 Preview is included in Kimi membership plans that carry Kimi Code: Plus and above on the new plans, Andante and above on legacy plans, per Moonshot's Kimi Code membership docs. Moonshot says the plan rename left prices unchanged, but the monthly prices render client-side on the plan page and are not confirmed in this record. After the plan quota runs out, an optional Extra Usage balance billed in RMB keeps requests going. Moonshot's [Open Platform](/hub/companies/moonshot-ai) pricing page carries no separate K2.8 line.
| Tier | Rate |
|---|---|
| Kimi membership, new plans (Plus and above) | kimi-for-coding is included from Plus upward; Go has no coding quota. See price notes for what is confirmed. |
| Kimi membership, legacy plans (Andante and above) | Legacy subscribers keep their tier names and quota rules; kimi-for-coding is included from Andante upward. |
Key Features
- 1M-token context window: Serves the full million-token window, four times the 262,144 tokens on the kimi-for-coding-highspeed route.
- Three-level adjustable reasoning: Supports reasoning_effort low, high and max (max by default), the same three levels K3 exposes inside Kimi Code.
- Drop-in upgrade at the same model ID: Served at the existing kimi-for-coding model ID, so no client changed anything to receive it.
- OpenAI- and Anthropic-compatible API: A Kimi Code Console API key works against 2 protocol base URLs, so third-party agents can call the model directly.
Pros
- Full 1M-token context at the kimi-for-coding model ID, with image and video input.
- Shipped as a no-config upgrade for anyone already calling kimi-for-coding.
- Usable from Claude Code and other agents through the Kimi Code API key, not only Moonshot's own clients.
Cons
- No SWE-bench, GPQA or other independently verified benchmark score has been published for this checkpoint.
- No per-token API option: usage counts against a Kimi membership quota, with Extra Usage billed in RMB after that.
- The model behind kimi-for-coding can change without a version bump, which makes pinned evaluations hard.
Frequently Asked Questions
How much does Kimi K2.8 Preview cost?
There is no per-token price: K2.8 Preview comes with a Kimi membership plan that includes Kimi Code, which means Plus or higher on the new plans and Andante or higher on legacy ones. Moonshot says plan prices did not change in the rename, but the figures render client-side and are not confirmed here.
How does Kimi K2.8 Preview perform on benchmarks compared to Kimi K3?
No benchmark score has been published for this checkpoint yet; Moonshot only says its performance is close to K3 with more efficient thinking than K2.7 Code. K3 itself reported 93.5% on GPQA Diamond in Moonshot's July 2026 tech blog, a K3 figure that should not be read as K2.8's.
Is Kimi K2.8 Preview open source?
No. Kimi K2.7 Code shipped open weights on Hugging Face under a Modified MIT license, but K2.8 Preview has no public weights or license. It is reachable only through Kimi Code, at the kimi-for-coding model ID, including from third-party tools via a Kimi Code API key.
Does Kimi K2.8 Preview train on user data?
Moonshot has not published a data-retention or training policy specific to this checkpoint, and no system card exists for it. Confirm retention and training-on-inputs behavior with Moonshot directly before sending proprietary or regulated code through the kimi-for-coding route.
Who should use Kimi K2.8 Preview, and who should avoid it?
It fits developers already on a Kimi plan with Kimi Code who want long-context coding help without a K3 quota bill. Teams that need a published benchmark, a pinned model version or per-token API billing are better served by K3, the open-weighted K2.7 Code, or a pay-as-you-go model.