ERNIE-4.5-VL-424B-A47B-Base

The model's architecture is extensively documented in the ERNIE 4.5 Technical Report (June 2025). It details a novel multimodal heterogeneous Mixture-of-Experts (MoE) framework, specifically identifying the use of a 630M parameter ViT encoder with adaptive-resolution and 2D RoPE. The report describes the 'modality-isolated routing' technique and the integration of 128 experts (64 text, 64 vision) with 8 active experts per modality. It also specifies the use of Grouped Query Attention (GQA) and the PaddlePaddle framework for training. While the base model is clearly a 'from-scratch' pre-trained foundation model, the specific initialization of the vision encoder versus the transformer backbone is well-defined.

Documentation mentions the model was trained on 'trillions of tokens' and incorporates 'high-quality Chinese textual and visual data' alongside global web data. While it describes a 'human-model-in-the-loop' iterative cycle for data refinement and mentions general categories (STEM, Chinese visual knowledge, web), it lacks a precise percentage breakdown of the dataset (e.g., exact ratios of code, web, books). Specific data sources are not named individually, and the filtering methodology is described in high-level conceptual terms rather than reproducible technical specifications.

The tokenizer is publicly accessible via the Hugging Face repository in both PaddlePaddle and PyTorch (Transformer-style) formats. The technical report and model cards confirm support for a 131,072 token context window. Vocabulary size and tokenization approach are verifiable through the provided `tokenizer_config.json` and `vocab.json` files in the official 'baidu/ERNIE-4.5-VL-424B-A47B-Base-PT' repository, ensuring alignment with the claimed multilingual (Chinese/English) support.

Baidu provides exemplary transparency regarding parameter counts for this MoE model. It explicitly states a total of 424 billion parameters with exactly 47 billion active parameters per token. The architectural breakdown is detailed, specifying 128 total experts with a 64/64 split between modalities and the activation of 8 experts per token. This prevents the common 'parameter inflation' ambiguity seen in other MoE releases.

The technical report mentions the use of 'heterogeneous hybrid parallelism' and 'FP8 mixed-precision training' on the PaddlePaddle framework. It discloses a Model FLOPs Utilization (MFU) of 47% for the largest language model variant. However, it fails to provide the total number of GPU/TPU hours, the specific hardware cluster size (e.g., number of H100 nodes used for the full pre-training run), or a calculated carbon footprint, which are requirements for a high score in this category.

The model provides results on standard benchmarks like MMMU, MathVista, and CV-Bench, and the technical report includes some hyperparameters (e.g., 2 FPS, 480 max frames for video). While evaluation code is partially available through the ERNIEKit and PaddlePaddle repositories, and inference is supported in vLLM, the exact prompts and few-shot examples used to achieve the reported SOTA scores are not fully disclosed in a centralized, reproducible format.

The model exhibits high identity consistency, correctly identifying as part of the ERNIE 4.5 family. Documentation clearly distinguishes between the 'Base' (pre-trained) and 'Chat' (post-trained) variants, as well as the 'Thinking' and 'Non-thinking' operational modes. Versioning is clear across the 10 distinct variants in the family, and there is no evidence of the model claiming to be a competitor's product.

The model weights, code, and development toolkits are released under the highly permissive Apache License 2.0. This is explicitly stated in the technical report, the GitHub repository, and the Hugging Face model cards. The license allows for both research and commercial use without the restrictive 'acceptable use policies' or 'research-only' clauses often found in other 'open' weights releases.

Hardware requirements are well-documented for various configurations. Official guides specify that ~945 GB of VRAM is needed for the 424B model, recommending 12x H100 (80GB) GPUs. It also provides clear guidance on 4-bit and 8-bit quantization (wint4/wint8) using the 'convolutional code quantization' algorithm, noting that 80GB x 8 resources are required for quantized deployment. Context length memory scaling is addressed through the mention of GQA and FlashAttention-like optimizations.

Baidu uses a clear naming convention (e.g., 4.5-VL-424B-A47B-Base) and maintains a GitHub repository for the ERNIE 4.5 family. However, there is no public, detailed changelog or semantic versioning system for weight updates (e.g., v4.5.1). While the release is recent, the infrastructure for tracking silent weight updates or behavior drift over time is not yet as robust as established open-source projects.