Key Takeaways
- Moonshot AI released Kimi K2.5 on January 27, 2026 — a 1-trillion-parameter Mixture-of-Experts model with native multimodal and agentic capabilities, available under a modified MIT license.
- Cursor built its Composer 2 coding model on top of Kimi K2.5 without initial attribution, sparking an industry controversy over open-source credit obligations in March 2026.
- The model scores 76.8% on SWE-Bench Verified and 50.2% on Humanity’s Last Exam, while activating only 32 billion of its 1 trillion total parameters per request.
- Moonshot AI’s valuation surged from $4.3 billion to an estimated $18 billion within three months of K2.5’s release, backed by Alibaba, Tencent, and 5Y Capital.
What Happened
Beijing-based Moonshot AI released Kimi K2.5 on January 27, 2026 — a native multimodal agentic model built through continual pretraining on approximately 15 trillion mixed visual and text tokens atop its predecessor, Kimi K2 Base. The model uses a Mixture-of-Experts (MoE) architecture with 1 trillion total parameters but activates only 32 billion per inference request, making it efficient enough to run on consumer-grade hardware with INT4 quantization.
The release initially attracted moderate attention in the open-source AI community. That changed in March 2026 when a developer discovered that Cursor’s new Composer 2 coding model was built on top of Kimi K2.5. An internal model identifier — accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast — exposed the dependency, and Cursor had not disclosed the relationship in its launch materials.
Why It Matters
Kimi K2.5 represents a shift in how open-source foundation models get adopted by commercial products. Cursor, valued at over $50 billion through its parent company Anysphere, used roughly a quarter of its Composer 2 pretraining from the Kimi K2.5 base model, with additional fine-tuning and reinforcement learning applied on top. Cursor co-founder Aman Sanger acknowledged the oversight: “It was a miss to not mention the Kimi base in our blog from the start. We’ll fix that for the next model.”
The controversy highlighted the growing importance of attribution in open-source AI. Kimi K2.5’s modified MIT license requires prominent credit for products exceeding 100 million monthly active users or $20 million in monthly revenue. Anysphere’s annual recurring revenue exceeds $2 billion, placing it well above that threshold. Moonshot publicly confirmed that Cursor accessed Kimi K2.5 through an authorized commercial partnership via Fireworks.
Technical Details
Kimi K2.5’s architecture comprises 61 transformer layers including 1 dense layer, with 384 experts selecting 8 per token. Each expert operates with a hidden dimension of 2,048 and uses SwiGLU activation. The model employs Multi-head Latent Attention (MLA) with a 7,168 attention hidden dimension, a 160K vocabulary, and a 256K context window. For vision tasks, the integrated MoonViT encoder adds approximately 400 million parameters, supporting images up to 4K resolution (4096×2160) and 2K video (2048×1080).
On benchmarks, K2.5 scores 76.8% on SWE-Bench Verified, 73.0% on SWE-Bench Multilingual, and 85.0% on LiveCodeBench. In vision understanding, it reaches 78.5% on MMMU Pro and 86.6% on VideoMMMU. On Humanity’s Last Exam, K2.5 achieves 50.2% at 76% lower cost than Claude Opus 4.5, and 74.9% on BrowseComp for web agent tasks.
The model’s Agent Swarm framework allows self-directed orchestration of up to 100 specialized sub-agents executing parallel workflows across up to 1,500 tool calls. Moonshot AI claims this parallel approach cuts execution time by 4.5x compared to sequential processing. Native INT4 weight-only quantization with group size 32 enables up to 2x generation speedups on Hopper-architecture GPUs.
Who’s Affected
AI-assisted coding tool developers face new scrutiny over model provenance. Cursor’s experience demonstrated that users and competitors will inspect upstream dependencies aggressively. Developers using Kimi K2.5 through the Moonshot API, OpenRouter, Together AI, or NVIDIA NIM now have access to a trillion-parameter model at a fraction of the cost of closed alternatives. Companies building on open-source models above the revenue or MAU thresholds must now evaluate attribution compliance more carefully.
For Moonshot AI, founded in March 2023 by Yang Zhilin, Zhou Xinyu, and Wu Yuxin — all Tsinghua University alumni — the K2.5 release and Cursor controversy accelerated commercial momentum. The company’s valuation trajectory went from $4.3 billion in January 2026 to a reported $18 billion by March, with backing from Alibaba, Tencent, and 5Y Capital.
What’s Next
Moonshot AI is reportedly weighing a Hong Kong IPO and seeking up to $1 billion in expanded funding. The Kimi K2.5 model weights remain available on Hugging Face for download and local deployment using vLLM, SGLang, or KTransformers. The Cursor attribution dispute is likely to influence how other companies — including those building on Meta’s Llama, Alibaba’s Qwen, and DeepSeek models — approach open-source licensing disclosure in commercial products.
