Kimi K2.5: A Trillion-Parameter Open Model You Can Actually Download

What Kimi K2.5 is

Kimi K2.5 is Moonshot AI's open-source native multimodal agentic model, and the headline number is unusual for something you can download: one trillion total parameters. It was not trained from scratch — Moonshot built it by continual pretraining on roughly 15 trillion mixed visual and text tokens on top of Kimi-K2-Base, which is why vision is native rather than bolted on. It runs in both instant and thinking modes, and is released under a modified MIT licence, so it can be self-hosted. It has a 256K token context window; Moonshot's later Kimi K3 has 1 million.

The architecture, which Moonshot publish in full

It is a mixture-of-experts model: one trillion total parameters with 32 billion activated per token, drawn from 384 experts of which 8 are selected per token alongside 1 shared expert. There are 61 layers, only one of them dense. Attention uses MLA across 64 heads with a hidden dimension of 7168, the MoE hidden dimension is 2048 per expert, and the activation function is SwiGLU. Vision is handled by MoonViT, a 400 million parameter encoder. The vocabulary is 160K.

How Moonshot benchmark it

The card carries a comparison table against the frontier, which is the vendor's own. On SWE-Bench Multilingual it records 73.0 against GPT-5.2 at xhigh effort on 72.0, DeepSeek V3.2 in thinking mode on 70.2 and Gemini 3 Pro at high thinking on 65.0, with Claude 4.5 Opus in extended thinking ahead of all of them at 77.5. Every competitor figure there is taken at a raised effort setting. Independently recorded elsewhere on this site: SWE-bench Verified at 74, OSWorld at 63, FictionLiveBench at 86, SimpleBench at 47 and SimpleQA Verified at 34.

LMArena

On LMArena's text leaderboard Kimi K2.5 has 1450 Elo (rank 82, thinking effort); by category, 1502 Elo in Coding (rank 79, thinking effort), 1470 Elo in Math (rank 55, thinking effort), 1482 Elo in Expert (rank 74, thinking effort) and 1422 Elo in Creative Writing (rank 83, thinking effort).

Price

Via third-party APIs it is $0.60 per million input tokens and $3 output, an aggregated figure rather than first-party pricing.

What Moonshot added over K2

Moonshot list three capabilities for K2.5: native multimodality from pre-training on vision and language tokens, covering visual knowledge, cross-modal reasoning and tool use grounded in visual input; coding with vision, generating code from UI designs and video workflows; and Agent Swarm, which decomposes a complex task into parallel sub-tasks run by dynamically created, domain-specific agents. It runs in instant and thinking modes and in conversational and agentic settings. A changelog entry dated 29 January 2026 records the removal of the default system prompt and a correction of the <|media_start|> token to <|media_begin|> in the chat template. The paper is published on arXiv as 2602.02276.

More of Moonshot’s table

In thinking mode Moonshot report HLE-Full 30.1 (50.2 with tools), AIME 2025 96.1, HMMT February 2025 95.4, IMO-AnswerBench 81.8, GPQA Diamond 87.6 and MMLU-Pro 87.1, against GPT-5.2 at xhigh effort on 34.5, 100, 99.4, 86.3, 92.4 and 86.7. On image and video benchmarks it records MMMU-Pro 78.5, MathVision 84.2 and MathVista (mini) 90.1. These are Moonshot's figures.

Official Resources