ERNIE-4.5-VL-28B-A3B-Thinking

2025-11-11 · Publisher : Baidu · Domain : Artificial Intelligence

Baidu releases ERNIE-4.5-VL-28B-A3B-Thinking, a vision-language model with 28 billion parameters, 3 billion of them active, under the Apache 2.0 license, which reasons over images and fits on a single 80 GB GPU.

### Context ERNIE-4.5-VL-28B-A3B-Thinking extends the ERNIE 4.5 family opened in June 2025. It improves the visual reasoning of the ERNIE-4.5-VL-28B-A3B variant through a long mid-training phase and multimodal reinforcement learning. Baidu compares it to leading closed models (GPT-5 and Gemini 2.5 Pro according to press coverage). It is also one of the first open models of this size to combine long reasoning, visual grounding and tool use in a single model, which brings it closer to the needs of document agents: reading charts, plans or forms, then acting on the result. ### What the model brings - Multi-step visual reasoning: chart analysis, causal and STEM reasoning from photos of problem statements, thanks to reinforcement learning on verifiable tasks (GSPO and IcePop to stabilize MoE training, with dynamic difficulty-based sampling). - “Thinking with images”: the model can zoom into an image and search for images via tools to resolve an ambiguity. - Reinforced visual grounding and better instruction following for industrial scenarios. - Tool calling and image search to identify long-tail knowledge. - Video understanding: spotting events and changes across different time segments. The card stresses temporal awareness and event localization in videos. ### Architecture and specifications - Architecture: vision-language MoE with 28 billion parameters in total, 3 billion active, derived from ERNIE-4.5-VL-28B-A3B. - Modalities: text, image, video. Languages: English and Chinese. Context and knowledge cutoff: not disclosed in the card consulted. - License: Apache 2.0 (free commercial use). - Weights: baidu/ERNIE-4.5-VL-28B-A3B-Thinking on Hugging Face; the card requires transformers 4.57.6 at most. ### Cost and access No API pricing was consulted. VentureBeat indicates that a single 80 GB GPU is enough, hardware it prices between $10,000 and $30,000 depending on the model. | Item | Information | | --- | --- | | Official Baidu API | Pricing not consulted | | Self-hosting | 1 80 GB GPU (VentureBeat, secondary source) | | License | Apache 2.0 | ### Benchmarks | Indicator | Information | | --- | --- | | Published comparison | Image benchmark.jpg in the Hugging Face card, without values in text | | Comparison models cited by the press | GPT-5, Gemini 2.5 Pro | | Independent validation | None at the date of the coverage consulted | Source: Hugging Face card and VentureBeat article. No numerical table usable in text: the scores are in an image in the card and no third-party measurement was found. ### Versus closed models No numerical score can be cited in text: Baidu states in its card that the model, with 3 billion active parameters, closely matches the industry's flagship models on several benchmarks, and VentureBeat reports the claim that it outperforms Google and OpenAI models on vision tasks (documents, charts, visual reasoning), noting that independent verification is lacking. These judgments are vendor claims, to be confirmed by third-party rankings. The price gap cannot be quantified. What is established is practical: a model under Apache 2.0 that fits on an 80 GB GPU makes it possible to process documents and charts without sending data to an external provider. ### Key takeaways ERNIE-4.5-VL-28B-A3B-Thinking brings visual reasoning with image zoom and tool calling in a lightweight model under Apache 2.0. The claims of parity with closed models remain to be verified by independent measurements.

Tags : LLM, Open Source, Baidu, ERNIE, Apache 2.0, MoE, Vision-langage