ERNIE-4.5-VL-28B-A3B-Base

The model is explicitly identified as a multimodal Mixture-of-Experts (MoE) transformer within the ERNIE 4.5 family. Baidu provides a technical report and GitHub documentation detailing a 'heterogeneous MoE' structure that uses modality-isolated routing to separate visual and textual processing. It specifies the use of Grouped-Query Attention (GQA), Rotary Position Embeddings (RoPE), and a variable-resolution Vision Transformer (ViT) encoder. While the high-level architecture is well-documented, specific layer-by-layer configurations and the exact pre-training data mixture are not fully disclosed.

Baidu mentions the use of a 'vast and highly diverse corpus' of visual-language reasoning data and 'premium' datasets during a mid-training phase. However, there is no detailed breakdown of the data sources (e.g., specific web crawls, book datasets, or code repositories) or the exact proportions of each modality. The filtering and cleaning methodologies are described in general terms ('systematic data construction') without providing verifiable metrics or access to sample data.

The model uses a tokenizer compatible with the PaddlePaddle and Transformers frameworks, with vocabulary and implementation details available through the official ERNIEKit and Hugging Face repositories. It supports a context length of 131,072 tokens. While the tokenizer's code is public, detailed documentation on its specific training data alignment and normalization procedures is less comprehensive than the architectural details.

Baidu is highly transparent regarding the MoE parameter distribution, explicitly stating a total of 28 billion parameters with 3 billion active parameters per token during inference. The documentation distinguishes between shared experts and dedicated modality-specific experts. This level of detail regarding sparse activation is exemplary compared to many competitors who only disclose total counts.

Documentation confirms the use of the PaddlePaddle framework and mentions optimizations for NVIDIA Hopper (FP8) and Ampere (INT8) architectures. However, specific compute metrics such as total GPU/TPU hours, the number of chips used, training duration, and the estimated carbon footprint are conspicuously absent from public reports.

Baidu provides results for standard benchmarks like MathVista, ChartQA, and OCRBench, and includes some evaluation scripts within the ERNIEKit repository. However, the exact prompts, few-shot examples, and specific versions for all benchmarks are not consistently detailed. Independent third-party verification is limited, and some results remain vendor-published without full reproduction instructions.

The model consistently identifies as part of the ERNIE 4.5 family and maintains clear versioning between its 'Thinking' and 'Base' variants. It is transparent about its multimodal nature and its specific 'thinking' vs 'non-thinking' operational modes. There are no documented instances of the model claiming to be a competitor's product or misrepresenting its core identity.

The model weights and associated code are released under the Apache License 2.0, which is a standard, permissive open-source license. This allows for both commercial and non-commercial use with clear terms. The license is prominently displayed on Hugging Face, GitHub, and in the technical report, with no conflicting proprietary terms found in the primary documentation.

Baidu provides specific VRAM requirements for different deployment scenarios, noting that 80GB is required for full FP16 inference on a single card, while quantization (WINT8) can reduce this to approximately 60GB. They also provide guidance for multi-GPU setups and vLLM integration. While helpful, more detailed scaling data for different context lengths and batch sizes would be required for a higher score.

The model follows a versioned release cycle (e.g., v1.0 to v1.5 of ERNIEKit) and maintains a changelog on GitHub. However, the documentation of 'silent' updates to the weights or changes in safety filtering is sparse. There is no formal system for tracking performance drift over time or a clear policy for accessing deprecated versions of the weights.