Moonshot AIKimi K2.5

External Link

Kimi-K2.5 is an open-weight vision-language model published by Moonshot AI.

Token context
262K tokens
Inputs
Text, Image
Outputs
Text
Released
Jan 27, 2026

Providers & Pricing

Estimate workload cost →
ProviderInput / 1MOutput / 1MCache read / 1MCache write / 1MObservedEvidence
$0.45$2.25$0.070—Aug 29, 2026Provider-reported
$0.50$2.50$0.15—Oct 19, 2026Provider-reported
LMArena Document Arenadocument-2026-09-13-a64ca3b12b6e1,430.10rating—21,569kimi-k2.5-thinking; 95% CI [1423.37027866, 1436.81983378]; votes 21569; rank 34unknownOriginal sourcewinner eligibleThird-party benchmark
LMArena Text Arenatext-2026-10-15-a64ca3b12b6e1,445.02rating—74,759kimi-k2.5-thinking; 95% CI [1441.75782956, 1448.27322629]; votes 74759; rank 62unknownOriginal sourcewinner eligibleThird-party benchmark
LMArena Vision Arenavision-2026-10-09-a64ca3b12b6e1,268.77rating—28,297kimi-k2.5-thinking; 95% CI [1263.09963340, 1274.44784290]; votes 28297; rank 44unknownOriginal sourcewinner eligibleThird-party benchmark

At a Glance

Model Facts

Identity

Status
active
Developer ID
moonshotai/Kimi-K2.5
Released
Jan 27, 2026
Version
Kimi-K2.5
License
other

Capacity

Token context
262K tokens
Maximum output
Unknown
Total parameters
1.03T
Active parameters
32B
Knowledge cutoff
Not reported

Interface and Access

Inputs
Text, Image
Outputs
Text
API available
Yes
Open weights
Yes
Self-hostable
Yes
Reasoning
Yes
Vision
Yes
Tool calling
Yes
Capabilities
chat, generation, reasoning, tools

This page represents one developer model product. Dated API snapshots, serving endpoints, pricing, regions, quantizations, and service tiers attach as versioned aliases or provider details and never create duplicate public model records.

Version Lineage

PredecessorUnknown
SuccessorsUnknown
Aliasesmoonshotai/Kimi-K2.5, moonshotai/Kimi-K2.5

Explicit Unknowns

  • Maximum output
  • Knowledge cutoff

Primary Evidence

Sources and Observation Date

Comparable Peers

Related Model Comparisons

All comparisons →
APairBContext
vscross-developer peersimage, text
vsfamily variantsembedding, image, text
vscross-developer peerstext
vsfamily variantstext
vsfamily variantsembedding, image, text
vsfamily variantsimage, text
vscross-developer peersimage, text

Questions

Kimi K2.5 FAQs

What is Kimi K2.5?+

Kimi-K2.5 is an open-weight vision-language model published by Moonshot AI. It is developed by Moonshot AI and its current sourced lifecycle status is active.

What inputs and outputs does Kimi K2.5 support?+

Kimi K2.5's current record lists Text and Image as input and Text as output.

How much context does Kimi K2.5 support?+

The current record lists a 262K-token context window and does not report a maximum output length.

Is Kimi K2.5 available through an API?+

Kimi K2.5 is marked as API-available. The directly sourced provider records currently include Deepinfra, Hugging Face, and Openrouter.

How large is Kimi K2.5?+

The current record lists 1.03T total parameters and 32B active parameters. Its architecture is KimiK25ForConditionalGeneration.

What capabilities and tasks does Kimi K2.5 support?+

Kimi K2.5's recorded capabilities are chat, generation, reasoning, and tools. Its supported tasks are embedding, image, and text.

How much does Kimi K2.5 cost through an API?+

The lowest directly sourced prices currently attached to Kimi K2.5 are $0.45 per million input tokens through Deepinfra and $2.25 per million output tokens through Deepinfra. Prices are provider-specific and should be checked against each cited observation date.

Can Kimi K2.5 be self-hosted?+

Kimi K2.5 is an open-weight model. Self-hosting is supported, and the recorded license is other.

Does Kimi K2.5 have related versions or aliases?+

No predecessor is recorded. No successor is recorded. Recorded aliases are moonshotai/Kimi-K2.5.

Which models can Kimi K2.5 be compared with?+

The current compatible peer set includes DeepSeek V4.1 Flash, Kimi K3, Qwen3.8 2.4T A95B, Kimi K2 Thinking, and Kimi K2.6. Compatibility requires shared sourced modalities and tasks; it does not imply that one model is better. Browse model comparisons →

Send Feedback