Skip to content
← Models
Open weightsMoonshot AI/10 trustVerified
Kimi K2.5 logo

Kimi K2.5

1T open-weight multimodal MoE with Agent Swarm: 100 parallel sub-agents.

Visit official page →

About

Kimi K2.5, released January 2026 by Moonshot AI, is a 1T-parameter MoE (32B active) built on natively multimodal training: 15T tokens of mixed text and vision from the start, with a 400M-parameter MoonViT vision encoder for images and video. It runs in Instant, Thinking, Agent, and Agent Swarm modes — the last letting it direct up to 100 sub-agents in parallel, cutting execution time ~4.5x. Moonshot claims 50.2% on Humanity's Last Exam at 76% lower cost than Claude Opus 4.5, and it open-sourced the weights under a modified MIT license.

The honest picture: 1T total parameters means serious infrastructure for self-hosting despite the sparse 32B active; most users go through the API or quantized builds. Vendor benchmark claims (beating GPT-5.2, Claude Opus 4.5) are self-reported and should be checked against independent evals. The agentic swarm features are genuinely novel but operationally complex — orchestrating 100 sub-agents needs careful prompt and cost management.

Specs

Context window

Pricing

Free quota

API access

Weights

Open weights

Released

January 27, 2026

Benchmarks

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

open-source-llmagentsmultimodalcoding

Trust score — 9/10

/10 trustVerified

Scored 9/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 46; Trending score 90/100 from researcher assessment.

More in the kimi-k2 family

Sources