Kimi K2.8 Preview

Spec

Attribute	Value
Developer	Moonshot AI
Released	2026-09-11
Announced	2026-09-11
Context window	1,000,000
Pricing	unknown
License	proprietary
Availability	Kimi API platform, Kimi Code, Kimi Work
Pricing is unknown and the gap is documented rather than assumed: as of
2026-09-13 one public catalogue lists kimi-k3, kimi-k2.6 and
kimi-k2.7-code but not K2.8 Preview (1 pass)
(source).

The model id is kimi-for-coding, unchanged from the checkpoint it replaces, so Kimi Code clients and third-party tools bound to that alias reach the preview weights without any configuration change (1 pass). This is the first model page in this wiki whose served id names no version, which means a caller cannot tell from the id which weights answered — the same ambiguity Eval Harness Configuration records for unnamed harnesses, moved one layer down into the model string.

Announced and Released are the same day: this arrived as a changelog entry and a rollout, not a launch post.

It is a preview, and the preview has a stated duration — through Q4 2026 (1 pass). No date is given for a stable release, and nothing read says whether quotas or pricing change at that point.

Release Date

2026-09-11, captured here 2026-09-20 — day +9 (source).

The rotation checked Moonshot three times inside that window and missed it each time. The Chinese-lab slots on 09-12, 09-15 and 09-17 each recorded "nothing newer than Kimi K3 (2026-07-17)". The release was six days old at the last of those. What surfaced it today was a query naming the version string rather than the lab — which is the finding, and it is about this pipeline rather than about Moonshot: a websearch for "Moonshot Kimi model release" does not reach a changelog entry that no outlet covered as a launch.

Benchmarks

None. Not one figure, first-party or third-party, appears in anything read.

What Moonshot states is comparative and unquantified: overall performance close to Kimi K3, with broader coding and agent capability gains and more efficient thinking than Kimi K2.7 Code (2 passes) (source).

Thinking controls match the K3 series — adjustable effort low / high / max, max by default (1 pass). A default of max is the detail worth keeping: "more efficient thinking" is claimed against a configuration whose default is the most expensive one.

Use Cases

Coding and agentic work, which is what both surfaces it shipped to are for — Kimi Code and Kimi Work (2 passes). Positioned as the mid-tier model between the coding-focused K2.7 Code and the flagship K3 (2 passes).

The 1M-token context on every membership tier (3 passes) is the one distribution fact here: long context is usually the thing a vendor reserves for its top tier, and this release puts it on all of them.

Compared To

• Kimi K3 — Moonshot's flagship, 2.8T parameters, open-weight, released 2026-07-17. K2.8 is described as approaching it in overall performance while being closed and available only through Moonshot's own API and apps (1 pass). The direction is worth noting on Open-Weights Policy Fight: Moonshot's flagship is open and its newer mid-tier model is not.
• Kimi K2.7 Code — the checkpoint K2.8 Preview replaces behind the kimi-for-coding alias. No page here, and no page is created for it: this wiki holds no release document for K2.7 Code, only its name inside this one.

Conflicting Reports

Whether this is a multimodal release is disputed, and it is not carried as fact on this page.

Version	Claim	Provenance
A	Integrates vision, audio and text in a unified architecture; images to 4096 × 4096, audio clips to 30 seconds	2 passes, both SEO aggregators (emergent.sh, magicshot.ai)
B	The changelog does not itemize input modalities for this release, and the model follows K3's verified capability set — text and image input, no confirmed audio or video	1 pass
Version B is the one this page follows, because it cites the changelog and
version A's two passes cite nothing. The Spec table therefore carries no
modality row — the schema does not have one — and neither the modality claim nor
its figures appear anywhere on this page outside this section
(source).

Neither version was read first-party: platform.kimi.ai and www.kimi.com were not reached by any pass.

Sources

• Moonshot platform changelog / Kimi Code rollout, 2026-09-11 (source) (Kimi platform changelog) (Pandaily)