Wait Which Model?
← Back to directory
Moonshot AI

Kimi K2.8 Preview

Released Sep 11, 2026 · knowledge cutoff unpublished

Status
Unknown
Location
China
Modality
Multimodal
Context window
1M
Max output
—
Speed
—
Price ($/MTok in / out)
— / —
Cost per task
—
Open weights
No
Benchmarks0 of 10 reported

MMLU-Proretired

—

GPQA Diamond

—

SWE-bench Verified

—

SWE-bench Pro

—

Terminal-Bench 2.1

—

AIME (latest)retired

—

Humanity's Last Exam

—

LMArena Elo

—

GDPval-AA v2

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • Takes screenshots and short videos of a broken UI directly in the coding thread — the first Kimi Code model that can, so a visual bug no longer has to be described in prose
  • Thinks more economically than K2.7 Code for what Moonshot calls near-K3 results — one early head-to-head finished an agent task in two-thirds of K3's wall-clock time and cost
  • Drops into existing setups with zero configuration: the kimi-for-coding alias was upgraded in place, and every paid tier now gets the full 1M window instead of rationing context

Weaknesses

  • Moonshot has published no benchmarks and no model card, so 'close to K3' is a vendor claim — early testers found it faster but not at K3's level, and it failed to one-shot a task K3 completed
  • Only reachable through Kimi Code and Kimi Work subscriptions — there is no per-token API price or open-platform listing, so it cannot be metered, pinned or self-hosted
  • The unchanged alias means the model underneath changed silently on 2026-09-11 — anything evaluated against K2.7 Code through kimi-for-coding is no longer reproducible

Subscription-only mid-tier model served in Kimi Code and Kimi Work from 2026-09-11 through the existing kimi-for-coding alias, which Moonshot upgraded in place from K2.7 Code (a model this site does not track, so predecessorId stays null even though the in-place replacement is explicit). Moonshot's docs give image and video input, a 1,048,576-token window on Plus and above (Andante and above on legacy plans) and low/high/max thinking effort — the models page says high is the default while the what's-new post says max. No benchmark table, model card, knowledge cutoff or per-token price exists: platform.kimi.ai lists K3 ($3/$15) and K2.7 Code ($0.95/$4.00) but no K2.8, and the $0.80-1 / $3.35-4 figures on third-party gateways are unofficial, so pricing is null and availability restricted. Not on Artificial Analysis, arena.ai or ARC Prize. Filed as balanced: Moonshot positions it between K3 and K2.7 Code for cost-effective everyday development.

For developers

API model strings

  • moonshotkimi-for-coding

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor

News