Overview
Kimi K2.8 Preview is Moonshot AI's coding and agent model, announced in the Kimi Code "What's New" changelog on 11 September 2026 as fully rolled out inside Kimi Code. The rollout kept the model ID unchanged at kimi-for-coding, so clients and third-party tools already pointed at that alias pick the new model up without a configuration change.
Moonshot positions it against its own two neighbours rather than against the field: performance is described as close to K3, with coding and agent capabilities "improved across the board" and "significantly more efficient thinking than K2.7 Code". No benchmark table accompanied the changelog entry, so this page carries none.
Two controls came across from the K3 series. Thinking effort is adjustable at low, high or max, with max the default. And when thinking is turned off, the changelog says requests to the K3 series and to K2.8 Preview are both served by K2.8 Preview in its no-thinking mode — so the preview model is also the fast path for the flagship line.
Context is the other headline: up to 1M tokens, and Moonshot states that window is available across all membership tiers rather than only the top plan. The model is a Kimi Code product release — it does not appear in the Kimi open API platform's model list, which documents kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed and kimi-k2.6, and no open weights have been published for it.
| Released | 2026-09-11 |
|---|---|
| License | Proprietary |
| Weights | API only |
| Context | Up to 1M tokens, across all membership tiers |
| Modalities | Text |
| Status | Preview — rolled out inside Kimi Code; not listed on the Kimi open API platform and no weights published |
Strengths
- 1M-token context window available on every Kimi Code membership tier, per Moonshot's changelog
- Drop-in for existing setups: the model ID stays kimi-for-coding, so no client reconfiguration is needed
- Three thinking-effort levels (low / high / max, default max) matching the K3 series
- Also serves K3-series requests when thinking is switched off, giving the line a cheaper no-thinking path
- Moonshot reports coding and agent gains over K2.7 Code with markedly more efficient thinking
Best for
- Long-horizon coding sessions in Kimi Code where a whole repository plus history has to stay in context
- Agentic tool-use runs that need effort dialled down for routine steps and up for hard ones
- Teams already wired to the kimi-for-coding alias who want the newer model without touching config
- Latency-sensitive work that can run with thinking turned off
How to access
| Provider | Model ID |
|---|---|
| Kimi Code ↗ | kimi-for-coding |
Kimi K2 — every version
The full lineage of the Kimi K2 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Kimi K2.8 Preview | 2026-09-11 | 1M | Proprietary |
| Kimi K2.7-Codecurrent | 2026-06-12 | 256K | Modified MIT |
| Kimi K2.6 | 2026-04-20 | — | Open weights |
| Kimi K2.5 | 2026-01-27 | — | Open weights |
| Kimi K2-Instruct-0905 | 2025-09-09 | — | Open weights |
| Kimi K2 | 2025-07-11 | — | MIT |
FAQ
Which model ID does Kimi K2.8 Preview use?
kimi-for-coding — unchanged from K2.7 Code. Moonshot's changelog makes the point explicitly: "The Model ID is unchanged — still kimi-for-coding", so Kimi Code and third-party tools wired to that alias need no configuration change.
How big is the context window?
Up to 1M tokens. Moonshot states the 1M window is available across all membership tiers, not just the highest plan.
How does it compare with Kimi K3?
Moonshot describes overall performance as close to K3 with more efficient thinking, and reuses K3's low / high / max effort levels (default max). With thinking turned off, requests to the K3 series are served by K2.8 Preview in its no-thinking mode. No benchmark figures were published with the announcement.
Are the weights open?
No. Unlike Kimi K2, K2.5, K2.6, K2.7-Code and K3, no weights have been published for K2.8 Preview, and it is not listed on the Kimi open API platform — it is served inside Kimi Code.