kimi-k2.5:cloud

Upstream reference: Kimi K2.5

Publisher: Moonshot AI

Moonshot AI describes a multimodal derivative of Kimi-K2-Base trained on mixed visual and text tokens, with instant and thinking generation modes.

Publisher-described context, separate from Sigma5C observations and recorded catalog statistics.

Published upstream specifications

Total parameters

1 trillion Source

Active parameters

32 billion Source

Published context

256K tokens, publisher notation Source

Vision encoder

MoonViT, 400 million parameters Source

Upstream license metadata

Modified MIT; review linked terms Source

Publisher-described uses

• Vision-supported coding, text generation and tool-oriented applications. Source

Reference limitations

• The model card’s multi-agent examples require orchestration around the model; a cloud alias is not a verified deployed digest. Source

A recorded tag and an upstream card association do not verify the deployed checkpoint, quantization, configured context or current hosted availability.

Publisher sources

• Kimi K2.5 official documentation — Moonshot AI. Checked 2026-11-02.

Links

Recorded references do not establish downloadable weights, licensing or current availability.

• Moonshot AI — official site
• Ollama
• Hugging Face
• GitHub (maker)

Where kimi-k2.5:cloud comes from

Named in 4 observation pages

• 31 cloud models × 3 wrappers on the same audit · V1
• Academic observation — 32 cloud models × Cline and Pi · V1
• Cloud Models — three-judge autoresearch observation set · V1
• Function-calling — 9 cloud models × 30 prompts · V1