Kimi K2.5 (TensorX)

Moonshot AI released Kimi K2.5 on January 27, 2026, and first posted its technical paper to arXiv on February 2, 2026; arXiv records show the paper was revised on August 7, 2026, making the current event a technical-paper update rather than a new model launch. The article explicitly identifies K2.5 as an open, native m The same source details K2.5's architecture as a 1-trillion-parameter mixture-of-experts model with 32 billion active parameters, and notes that Kimi's Agent Swarm feature can coordinate up to 100 sub-agents with claimed elapsed-time reductions of as much as 4.5x. These are the key K2.5-specific technical details suppo

Third-party model card attributes Kimi K2.5 to Moonshot AI as an open-weights release dated January 2026, configured as a 1040B-parameter mixture-of-experts model with 32B active parameters and a 256K-token context window. The card lists developer-reported benchmark scores including SWE-bench Verified at 76.8%, Terminal-Bench 2.0 at 43.2%, LiveCodeBench at 85%, GPQA Diamond at 87.6%, AIME 2025 at 96.1%, and MMLU-Pro at 87.1%, framing these as the published standard for first-party measurement with confidence intervals. The page situates Kimi K2.5 alongside comparable frontier models such as Gemini 3.1 Pro, DeepSeek-V4-Pro, GPT-5.5, Claude Opus 4.8, and Llama 4 Maverick for head-to-head evaluation, and provides hardware-sizing estimates across consumer and professional GPUs ranging from IQ3 at roughly 403GB to FP16 at 2080GB. All listed numbers are labelled developer-reported rather than independently reproduced on the page itself.

Moonshot AI positions Kimi K2.5 as an open-source multimodal model that turns text, image, and video inputs into production-ready frontend code, with an Agent Cluster mode that coordinates up to 100 sub-agents in parallel and claims up to a 4.5x speedup on long-horizon workflows. The model is available through Kimi Web, App, API, and Kimi Code, exposing four runtime modes: Instant for quick answers, Thinking for complex questions, Agent for research and content creation, and Agent Cluster (Beta) for large-scale tasks. K2.5 also powers dedicated Kimi agents for documents (Word and LaTeX-formatted PDFs with inline annotations), presentations (structured, downloadable slide decks), Excel (formula, pivot table, and chart generation with live data linkage), and vibe-coded website publishing from natural language and design references. A separate deep research workflow pulls from multiple sources and synthesizes them into visualized reports with charts and structured conclusions for investment, industry, academic, and strategy use cases.

Moonshot AI announced via its official Kimi Weibo account that Kimi K2.5, described as the company's first trillion-parameter multimodal model, will be retired at the end of August 2026. The model had been launched and open-sourced in January 2026 and featured a native multimodal architecture supporting both visual and The retirement marks Moonshot AI's official shift of focus to its successor, Kimi K3, which was open-sourced in July 2026 and is described as the first open-source model to reach 2.8 trillion parameters. K3 uses a KDA hybrid linear attention mechanism combined with attention residual technology and natively supports vi