Open-source base model on which Composer 2 and 2.5 are built. Cursor drew criticism for under-disclosing this at first.

Kimi K2.5
by Moonshot AI
Moonshot AI's open-weights 1T-parameter MoE multimodal agentic model with 256K context, vision input and agent swarm.
Updated EU status checked Last shown
Kimi K2.5 at a glance
- What is Kimi K2.5?
- Kimi K2.5 is a model from Moonshot AI. Moonshot AI's open-weights 1T-parameter MoE multimodal agentic model with 256K context, vision input and agent swarm.
- Who makes Kimi K2.5?
- Kimi K2.5 is made by Moonshot AI.
- Does Kimi K2.5 work in the EU?
- Available in the EU. Weights are downloadable on Hugging Face under a modified MIT license with no EU exclusion in the provided material, and the model is tagged for SageMaker deployment. Moonshot's hosted services and their EU data handling were not verified. Last verified 14 Nov 2026.
- How much does Kimi K2.5 cost?
- Open source. Weights free to download; Moonshot also offers paid API access and the Kimi chat app (pricing not in provided material). Licence: Modified MIT.
- Where was Kimi K2.5 shown?
- In 1 creator video by Matthew Berman. Most recently by Matthew Berman on 26 May 2026, at 7:13 in “Cursor just beat EVERYONE.”.
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Kimi K2.5 is a native multimodal mixture-of-experts model from Moonshot AI, built by continual pretraining on roughly 15 trillion mixed vision and text tokens on top of Kimi-K2-Base. It has 1 trillion total parameters with 32 billion active per token, 384 experts, a 256K-token context window, and a 400M-parameter MoonViT vision encoder.
It supports instant and thinking modes and is aimed at agentic use. Moonshot highlights coding from visual specifications such as UI designs and video workflows, and an "Agent Swarm" approach in which a task is split into parallel sub-tasks run by dynamically created agents. Cursor's Composer 2 and 2.5 models are reportedly built on it, according to a YouTube creator.
Weights are published on Hugging Face under a modified MIT license, in compressed-tensors format with custom code for transformers. The model is also available through Moonshot's API platform and the Kimi chat product. Running it locally needs substantial hardware because of its size. Moonshot's site now also lists newer Kimi K3 and K3.1 Pro models.
- 1T total / 32B active parameter MoE architecture
- Native multimodal input (text and images) via MoonViT encoder
- 256K-token context window
- Instant and thinking modes
- Agent Swarm parallel sub-agent execution
- Open weights under modified MIT license
As seen on YouTube
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Similar AI tools
Closest neighbours by meaning, not just by category.
Kimi K3.1
Model · Moonshot AI
Moonshot AI's open-weight 2.8T-parameter multimodal MoE agentic model with 1M-token context.
Kimi K3
Model · Moonshot AI
Moonshot AI's Kimi K3 model family, with K3.1 Pro as the current release in the Kimi app and API.
Kimi 3 Max
Model · Moonshot AI
Moonshot AI's large language model, described as open-source with roughly 2T+ parameters, benchmarked by Artificial Analysis.
Kimi K2.7 Code
Model · Moonshot AI
Moonshot AI's Kimi K2.7 coding model, offered through the Kimi Code agent (CLI, VS Code extension, desktop app).
LongCat 2.5
Model · Meituan
Meituan's open-weights 1.6T-parameter multimodal MoE model (48B active) for agentic coding and visual tasks.
MiniMax M3
Model · MiniMax
Open-weights native multimodal MoE model from MiniMax, ~428B parameters (~23B active) with 1M-token context.