Kimi K2.5

Kimi K2.5, released on 27 January 2026, is the release in which Moonshot AI's open trillion-parameter line became natively multimodal. It was not trained from scratch: the company continued pretraining Kimi K2 Base on roughly 15 trillion mixed visual and text tokens, so images enter the model directly through a 400-million-parameter MoonViT encoder instead of arriving as somebody else's caption. The architecture is otherwise the K2 body — mixture of experts with one trillion total parameters and 32 billion active per token, 61 layers, 384 experts of which eight are selected, MLA attention, a 160K vocabulary — with the context window doubled to 256K tokens. The headline idea is the agent swarm: for a complex task the model can create and orchestrate up to 100 sub-agents on its own, running up to 1,500 tool calls in parallel with no predefined workflow, which Moonshot says cuts execution time up to 4.5-fold. The swarm is measured, not just described: BrowseComp rises from 60.6 for the plain model to 74.9 with simple context management and 78.4 in swarm mode, and WideSearch from 72.7 to 79.0. Other vendor figures are 30.1 on HLE-Full and 50.2 with tools, 96.1 on AIME 2025, 87.6 on GPQA-Diamond, 76.8 on SWE-Bench Verified, 85.0 on LiveCodeBench v6 and 92.3 on OCRBench. As with K2 Thinking, Moonshot blocked the model's access to Hugging Face during the HLE run to rule out test-data leakage. One practical point a reader should know: the weights remain open under a modified MIT licence, but Moonshot's own API price list no longer carries K2.5 at all — it lists only K3, K2.7 Code and K2.6 — so the model now lives through self-hosting and third-party providers rather than through its maker's service.

▸Videos

No videos yet.

▸Reviews