OpenBMB released Trident-Qwen3VL-2B, a multimodal embedding model for mixed-modality retrieval
OpenBMB released Trident-Qwen3VL-2B on Hugging Face, a multimodal embedding model fine-tuned from Qwen3-VL-2B-Instruct on Trident-Train. It maps text, images, and fused inputs into one shared space for mixed-modality retrieval. The model is available under the MIT License.
China context
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download the weights from Hugging Face and integrate the model for mixed-modality retrieval tasks, subject to the MIT license for Trident-specific contributions and the Apache-2.0 license of the base model.
- For investors
- The release of an open-weights multimodal embedding model by OpenBMB signals continued investment in retrieval infrastructure, but no commercial terms or enterprise support are disclosed in the evidence.
OpenBMB released Trident-Qwen3VL-2B, a multimodal embedding model for mixed-modality retrieval, on Hugging Face. The model is fine-tuned from Qwen3-VL-2B-Instruct on the Trident-Train dataset and maps text, images, and fused text-image inputs into a shared embedding space, enabling a single text query to rank text-only, image-only, and fused documents. The LoRA adapter is already merged, so no base model or PEFT is needed. The Trident-specific contributions are released under the MIT License, while the base model Qwen3-VL-2B-Instruct is under Apache-2.0 and training images remain subject to their original licenses.
Trident-Qwen3VL-2B is a multimodal embedding model that unifies text, image, and fused inputs in a single embedding space. It is fine-tuned from Qwen3-VL-2B-Instruct using LoRA, with the adapter merged into the weights. The model uses different default instructions for queries and documents as part of its training format. The associated paper is 'Chaos in the Text: Revealing the Modality Preference in Mixed-Modality Retrievers' (arXiv:2610.11816).
Developers building retrieval systems for mixed text-image corpora can now use an open-weights model that ranks text, image, and fused documents in a single list, reducing the need to maintain separate embedding pipelines. The MIT license for Trident-specific weights lowers adoption friction, though the underlying Qwen3-VL-2B-Instruct remains Apache-2.0 and training images carry their own licenses.
Organizations with multimodal search or retrieval needs can evaluate Trident-Qwen3VL-2B as a drop-in embedding model that handles text, image, and fused queries without separate infrastructure. The MIT license on Trident-specific weights permits commercial use, subject to the base model and training data licenses.
The next verifiable signal is whether OpenBMB releases evaluation results on the Trident-Eval dataset or third-party benchmarks for mixed-modality retrieval. Another signal is community adoption, such as forks or integrations of the Trident repository on GitHub.