The Yi-9B model is explicitly documented as a depth-upscaled version of the Yi-6B base model. The technical report and Hugging Face documentation detail its decoder-only Transformer architecture, which is compatible with the Llama framework. It specifies the use of Grouped-Query Attention (GQA), SwiGLU activation functions, Rotary Position Embedding (RoPE), and RMSNorm. The training methodology, involving architectural expansion followed by 0.8 trillion tokens of incremental pre-training, is clearly described in the official 'Yi: Open Foundation Models' paper.

01.AI discloses that the model was trained on a total of 3.9 trillion tokens (3.1T for the base Yi-6B plus 0.8T for the 9B expansion). While they provide a high-level breakdown of the 3.1T corpus (English and Chinese web data, books, and code) and describe a cascaded filtering pipeline (MinHash deduplication, heuristic rules, and learned filters), they do not provide specific percentage breakdowns or public access to the raw datasets. The 0.8T incremental data is described generally as 'high-quality' with a focus on math and code.

The tokenizer is publicly available via Hugging Face and the official GitHub repository. It uses Byte-Pair Encoding (BPE) via the SentencePiece framework with a documented vocabulary size of 64,000 tokens. Technical details, such as splitting numbers into individual digits and handling rare characters via unicode-byte fallback, are explicitly stated in the technical report. The tokenizer is fully compatible with standard libraries like Hugging Face Transformers.

The model's parameter count is precisely stated as 8.83 billion (commonly rounded to 9B). As a dense model, all parameters are active during inference. The technical report provides an architectural breakdown, including the number of layers (44) and the specific attention mechanism (GQA). The impact of depth-upscaling on parameter density is well-documented in the context of its evolution from the 6B variant.

While the technical report mentions the use of a 'scalable super-computing infrastructure' and cross-cloud elastic task scheduling, it lacks specific details on the total GPU/TPU hours consumed, the exact hardware models used for the 9B training phase, or the associated carbon footprint. There is no detailed breakdown of the compute costs or environmental impact, which are key requirements for a high score in this category.

01.AI provides results for standard benchmarks like MMLU, C-Eval, and GSM8K in their technical report. However, the exact prompts and few-shot examples used for these evaluations are not fully disclosed in a reproducible format. While they mention using zero-shot and few-shot methods, the lack of a public evaluation harness or specific versioning for all benchmarks limits third-party verification. (Score adjusted due to undisclosed contamination risks).

The model consistently identifies itself as part of the Yi series developed by 01.AI. It maintains clear versioning (e.g., Yi-9B vs Yi-9B-200K) and is transparent about its bilingual focus and technical reasoning capabilities. There are no documented instances of the model claiming to be a competitor's product or misrepresenting its fundamental nature as an AI.

The model weights are released under the 'Yi Series Models Community License Agreement,' which is a custom license. While it allows for free commercial use for many users, it includes restrictions based on user scale (requiring explicit permission for large-scale commercial use) and specific regional legal compliance (PRC laws). This complexity and the distinction between the Apache 2.0 code license and the restrictive weight license create moderate ambiguity for users.

01.AI provides clear hardware requirements on their Hugging Face and GitHub pages, including minimum VRAM estimates for different model sizes. They explicitly support and document quantization methods like GPTQ and AWQ, providing scripts for users to perform these optimizations. VRAM requirements for 4-bit and 8-bit versions are generally well-communicated, though detailed context-length memory scaling curves are less prominent.

The model follows a basic versioning scheme (e.g., Yi-9B, Yi-1.5-9B), and 01.AI maintains a GitHub repository with some update history. However, there is no formal, detailed changelog or semantic versioning system that tracks minor weight updates or behavioral drift over time. Users must rely on date-stamped announcements rather than a structured versioning protocol.