Kimi-Dev-72B

The model's base architecture is explicitly identified as Qwen2.5-72B. Moonshot AI provides detailed documentation on the architectural modifications, specifically the 'duo-stage' framework consisting of 'BugFixer' and 'TestWriter' behaviors. The training methodology is well-described, involving a multi-stage process: mid-training on 150B tokens followed by large-scale reinforcement learning (RL) using outcome-based rewards in Docker environments. Technical reports and GitHub documentation provide a clear lineage from the base model to the specialized variant.

Moonshot AI discloses the general composition of the mid-training dataset (~150B tokens) as being sourced from 'millions of GitHub issues and PR commits.' While the total token count and general categories are provided, there is no granular breakdown of the specific proportions of different programming languages or repository types. The filtering and cleaning methodology is mentioned ('strict data decontamination' to exclude SWE-bench repositories), but the exact criteria for 'high-quality' data remain proprietary and lack public verification.

The model utilizes the standard Qwen2 tokenizer, which is publicly available and well-documented. The vocabulary size (151,936 tokens) and tokenization approach (BPE) are clearly stated in the configuration files on Hugging Face. The tokenizer's alignment with the claimed 131,072-token context window is verified through official vLLM serving instructions and model configuration files.

The model is explicitly described as a dense 72-billion-parameter architecture. Unlike MoE models, there is no ambiguity regarding active vs. total parameters. The parameter count is consistent across all official sources (Hugging Face, GitHub, and technical reports). The model configuration files provide a transparent breakdown of the architecture, confirming it is a standard dense transformer model.

Information regarding training compute is extremely limited. While the use of a 'highly parallel, robust, and efficient internal agent infrastructure' is mentioned, there are no specific disclosures regarding GPU/TPU hours, hardware specifications used for training, or the total training duration. Furthermore, there is no calculation of the carbon footprint or estimated compute cost, which are key requirements for high transparency in this category.

The model's performance on SWE-bench Verified (60.4%) is prominently featured with supporting documentation. Moonshot AI provides evaluation scripts and example result files on GitHub to facilitate reproduction. However, the exact prompts and few-shot examples used for all benchmarks are not fully disclosed in a centralized manner, and third-party verification is currently limited to community-run leaderboards rather than formal audits.

Kimi-Dev-72B demonstrates high identity consistency, correctly identifying itself and its specialized purpose in software engineering. It does not exhibit confusion with its base model (Qwen) in its primary documentation and maintains a clear versioning distinction from other models in the Kimi family (e.g., Kimi K2). The model's capabilities and limitations regarding autonomous agentic behavior are transparently communicated.

The model is released under the MIT License, which is a clear, permissive open-source license. The documentation explicitly states that the weights and source code are open for research and commercial use. There is a slight complexity noted regarding the upstream Qwen License Agreement, but the terms for Kimi-Dev-72B itself are clearly articulated in the LICENSE.md file on both GitHub and Hugging Face.

Moonshot AI and community documentation provide specific VRAM requirements for various deployment scenarios (e.g., ~145GB for FP16, ~65GB for quantized versions). Recommended hardware configurations (e.g., 8x RTX 4090 or 2x A100/H100) are publicly available. While the official documentation provides good baseline requirements, detailed memory scaling charts for the full 128K context length are less comprehensive.

The model uses a clear naming convention (Kimi-Dev-72B), but there is no evidence of a formal semantic versioning system or a detailed public changelog for weight updates. While the release date is clear, the lack of a structured path for tracking future iterations or documenting silent performance drifts results in a lower score for this criterion.