Kimi K2.5 is extensively documented as a 1-trillion parameter Mixture-of-Experts (MoE) model built upon the Kimi-K2-Base. The technical report (arXiv:2602.02276) and official GitHub repository provide deep architectural specifics, including the use of Multi-head Latent Attention (MLA), a 384-expert MoE structure (8 active per token), and the MuonClip optimizer for training stability. The model's 'Agent Swarm' paradigm and Parallel Agent Reinforcement Learning (PARL) methodology are also publicly detailed.

While Moonshot AI discloses the scale of the training data (15 trillion mixed visual and text tokens) and the methodology of 'joint text-vision pre-training' from the start, specific data sources and the exact percentage breakdown of the dataset remain undisclosed. The documentation mentions 'diverse internet data' and 'multimodal files' but lacks the granular transparency required for a higher score, such as specific domain proportions or detailed filtering/cleaning logs.

The tokenizer is publicly accessible via the official GitHub and Hugging Face repositories. It features a vocabulary size of 160,000 tokens and uses a specialized configuration to support native multimodality (e.g., <|media_begin|> tokens). The alignment between the tokenizer and the claimed language/multimodal support is verifiable through the provided code and model cards.

Moonshot AI provides exemplary transparency regarding parameter density. They explicitly state the total parameter count (1T) and the active parameters per forward pass (32B). Furthermore, the architectural breakdown is highly detailed: 61 layers (including 1 dense layer), 384 experts, 8 selected experts per token, and a 400M parameter vision encoder (MoonViT). This level of detail is rare for models of this scale.

Despite the detailed architectural info, there is a significant lack of transparency regarding training compute. No specific GPU/TPU hours, hardware cluster specifications, or carbon footprint calculations are provided in the technical report or official documentation. The mention of the 'MuonClip' optimizer for efficiency does not substitute for concrete compute resource disclosure.

Moonshot AI provides comprehensive evaluation results across standard (AIME 2025, MMLU-Pro) and internal (Kimi Code Bench, AI Office Bench) benchmarks. They disclose evaluation settings (temperature 1.0, top_p 0.95) and provide a technical report. However, the use of 'internally developed evaluation frameworks' for key coding tasks and the lack of full evaluation code for all reported metrics limit complete third-party reproducibility.

The model demonstrates high identity consistency, correctly identifying itself as Kimi K2.5 and distinguishing between its operational modes (Instant, Thinking, Agent, Swarm). Documentation clearly outlines versioning (K2 vs K2.5) and capabilities. There are no reported instances of the model claiming to be a competitor's product or misrepresenting its AI nature.

The model is released under a 'Modified MIT License.' While it allows for commercial use, modification, and distribution, it includes a specific 'attribution' clause requiring products with >100M MAU or >$20M monthly revenue to prominently display 'Kimi K2.5' on the UI. This is a clear, albeit restrictive, license that is well-documented in the GitHub repository.

Hardware requirements are thoroughly documented for various deployment scenarios. Official guides specify VRAM needs for INT4 quantization (the native format) and provide recommended configurations for production (e.g., 8x H200/B200 nodes). Third-party documentation from Unsloth and KTransformers further details extreme compression requirements (24GB VRAM + 256GB RAM) for local execution.

Moonshot AI maintains a basic changelog on Hugging Face (e.g., fixing system prompt issues and token names). However, the model has shown signs of 'effective compression degradation' in long-context windows according to independent analysis, and there is no formal, detailed version history or public deprecation roadmap for older Kimi K2 variants.