Moonshot AI provides high transparency regarding the Kimi K2-Instruct architecture. It is explicitly identified as a Mixture-of-Experts (MoE) model with 1 trillion total parameters and 32 billion active parameters. Documentation details 61 layers, 384 experts (8 active per token), and the use of Multi-head Latent Attention (MLA) with 64 heads. The training methodology is well-documented, specifically citing the use of the proprietary MuonClip optimizer to maintain stability during the 15.5 trillion token pre-training phase. The model is clearly identified as a post-trained instruction-following variant of the Kimi-K2-Base model.

While the scale of the training data is disclosed (15.5 trillion tokens), the specific composition and sources remain largely opaque. Official documentation describes the data as 'diverse' and mentions 'simulated multi-step tool interactions' for agentic training, but lacks a granular breakdown of data types (e.g., percentage of code, web, or academic papers). There is no public disclosure of specific datasets used for pre-training or the exact filtering and cleaning methodologies employed, which is a significant gap in upstream transparency.

The tokenizer is publicly accessible via the official Hugging Face repository and integrated into standard libraries like vLLM and SGLang. Technical specifications are clearly stated, including a vocabulary size of 160,000 tokens and support for a 128,000 token context window (extended to 256k in later revisions). The tokenizer implementation has been updated to fix specific bugs related to special token encoding and multi-turn tool call templates, with these changes documented in the public changelog.

Transparency regarding parameter density is exemplary for an MoE model. Moonshot AI explicitly distinguishes between the 1 trillion total parameters and the 32 billion parameters activated per token. The documentation further breaks down the architecture into 384 experts, 8 selected experts per token, and 1 shared expert. Detailed dimensions for attention (7168) and MoE hidden layers (2048) are provided, ensuring no ambiguity regarding the model's sparse nature or active compute requirements.

Information regarding training compute is limited to hardware types and general scale. While it is known the model was trained on large-scale GPU clusters (NVIDIA H100/H800/B200 mentioned in deployment contexts), the specific number of GPU hours, total energy consumption, and carbon footprint are not disclosed. The documentation focuses on the stability of the training process rather than the environmental or resource costs associated with the 15.5T token training run.

Moonshot AI reports extensive benchmark results (MMLU: 89.5%, SWE-bench Verified: 65.8%, MATH-500: 97.4%) and provides some evaluation parameters, such as output token lengths. However, the full evaluation code and exact prompts used for all benchmarks are not fully public. While third-party verification from platforms like Artificial Analysis exists, the lack of a comprehensive, reproducible evaluation suite in the official repository prevents a higher score. (Score adjusted for disclosed contamination issues).

The model demonstrates strong identity consistency, correctly identifying itself as 'Kimi, an AI assistant created by Moonshot AI' in official system prompts and API documentation. It maintains clear versioning (e.g., v1.0, 0711-preview, 0905-revision) and distinguishes itself from its 'Thinking' and 'Base' variants. There are no documented instances of the model claiming to be a competitor's product or misrepresenting its core identity.

The model is released under a 'Modified MIT License.' While this permits broad use, including commercial application, it includes specific restrictions (e.g., attribution requirements for products with >100M users or >$20M monthly revenue). The license is clearly stated in the GitHub and Hugging Face repositories, but the 'modified' nature introduces some complexity compared to standard open-source licenses like Apache 2.0.

Hardware requirements are thoroughly documented for various deployment scenarios. Official guidance specifies the need for approximately 1.1 TB of VRAM for FP8 weights and provides minimum configurations for quantized versions (e.g., 2-bit requiring ~381 GB). The documentation includes specific GPU cluster recommendations (e.g., 16x H800 for 128k context) and integrates with multiple inference engines (vLLM, SGLang, KTransformers), offering clear guidance on VRAM and system RAM trade-offs.

Moonshot AI maintains a public changelog on Hugging Face, documenting updates such as chat template improvements and tokenizer bug fixes. Semantic-style versioning is used for major releases (K2, K2.5) and specific checkpoints (0711, 0905). However, the frequency of 'silent' updates to the API endpoints and the lack of detailed deprecation paths for older weights prevent a higher score in this category.