BAAI

Foundation VLA model for bimanual dexterous manipulation. Pre-trained on multi-source egocentric data from humans (via MANUS data gloves) and robots.…

Tactile Hierarchical Dynamic dataset with 2.4M+ contact samples for embodied AI. ToucHD-Force collects 722,436 touch-force pairs using five tactile se…

Large-scale native multimodal world model that predicts the next vision-language state, pre-trained end-to-end on over 10T interleaved vision-language…

All-atom generative world model for all-to-all biomolecular interaction design. Allows scientists to specify epitopes on arbitrary targets and generat…

"Uniform Discrete Diffusion with Metric Path for Video Generation." BAAI Vision's successor to NOVA, replacing non-quantized autoregression with unifo…

Contamination-free evaluation of large reasoning models (LRMs) across 10 text reasoning tasks and 8 visual reasoning tasks. ROME provides 281 image-qu…

Embedding model for reasoning-intensive retrieval tasks. Achieves nDCG@10 of 37.1 on the BRIGHT benchmark. Represents a new direction for BGE — embedd…

Ultra-long video understanding model using task-aware KV sparsification. Developed jointly with Shanghai Jiao Tong University. Uses SigLIP-SO400M visu…

Multi-agent framework for high-quality scientific survey generation using a reflect-when-you-write paradigm. Features six specialized agents (Interpre…

Flexible open-source framework for comprehensive multimodal model evaluation across vision-language understanding and generation tasks (VQA, text-to-i…

Next-generation embodied AI VLM supporting spatial understanding and temporal decision-making. RoboBrain 2.5 adds precise 3D spatial reasoning and den…

First 3D-aware reasoning VLM for multi-step spatial referring (NeurIPS 2025). Integrates a depth encoder via SFT and uses reinforcement fine-tuning (R…

First comprehensive safety benchmark for video LVLMs. 2,264 video-text pairs covering 13 unsafe categories and 48 fine-grained subcategories, each pai…

Code embedding model supporting natural language queries in English and Chinese plus 20 programming languages. Also maintains strong general text retr…

Large-scale bilingual pretraining dataset (~35 TB) engineered for superior data quality and diverse reasoning trajectories. Comprises CCI4.0-M2-Base (…

Open-source initiative to unite the global community in developing next-generation language models, inspired by DeepSeek. Uses DeepSeek V3 MoE archite…

First open-source embodied operating system for cross-embodiment and multi-agent robot collaboration. Uses a Brain-Cerebellum hierarchical architectur…

Open-source platform for biomolecular 3D structure prediction and conformational ensemble sampling. Part of BAAI's WuJie series focused on life scienc…

Open-source protein representation learning framework using self-supervised sequence-structure collaborative learning. Features a robust structural en…

State-of-the-art multimodal embedding models for visual search including text-to-image, image-to-text, image+prompt-to-image, and text-to-image+text r…

Unified brain model for robotic manipulation from abstract to concrete. CVPR 2025. Foundational embodied intelligence model in BAAI's WuJie series.

Non-quantized autoregressive video generation model (ICLR 2025). Reformulates video generation as non-quantized autoregressive frame-by-frame temporal…

2B-parameter vision-language model built on LLaVA-one-vision framework with Qwen2.5-1.5B-instruct as LLM and SigLIP-SO400M as vision tower. Trained on…

"Infinity-MM: Scaling Multimodal Performance with Large-Scale and High-Quality Instruction Data." Open multimodal instruction dataset with 40M+ sample…

Unified image generation model supporting text-to-image, subject-driven generation, identity-preserving generation, image editing, and image-condition…

Native multimodal model trained solely via next-token prediction, demonstrating that a single architecture can match task-specific methods across gene…

Medical domain language model built on the Aquila foundation through continued pre-training, supervised fine-tuning, and reinforcement learning. Capab…

End-to-end distributed training and inference framework for large models built on open-source projects (Megatron-LM, vLLM). Supports pooled training a…

Unified multilingual, multi-functional, multi-granularity embedding model. Supports 100+ languages and combines dense, sparse, and multi-vector retrie…

AI framework donated to the Linux Foundation.

Second-generation multimodal model from BAAI, advancing vision-language generation with improved architecture and training. Supports any-to-any genera…

Second generation of the Aquila bilingual LLM. Improved with EfficientScale training recipe. Released in 7B and 34B sizes, plus an 8x16B MoE variant.

BAAI's first open-source bilingual LLM, the first compliant with Chinese regulations. Released in 7B and 33B sizes with AquilaChat instruction-tuned v…

One of the world's most widely used open-source embedding models for RAG. Includes BGE-v1.0 and BGE-v1.5 text embedding variants, plus BGE-Reranker fo…

Official library for the BGE model family and the C-MTEB benchmark for Chinese text embeddings and retrieval evaluation.

BAAI's first large-scale multimodal foundation model for any-to-any generation. Pioneered multimodal generation at BAAI, combining vision and language…

Comprehensive benchmark for vision-language models.

State-of-the-art open-source alternative to OpenAI's CLIP.

Platform for standardized model benchmarking (including OpenSeer).

"In-context" visual learning model that performs visual tasks without retraining.

Bilingual (Chinese/English) models for multimodal representation and generation.

A hierarchical transformer-based text-to-image model (6B parameters) that introduced the Cross-modal General Language Model (CogLM) and Local Parallel…

A large-scale cost-effective pre-trained language model series (up to 198B parameters) that introduced Mixture-of-Experts (MoE) and multilingual capab…

Landmark 1.75 trillion parameter MoE model, at the time the largest in the world, outscaling GPT-3.

Massive 5TB high-quality Chinese/English dataset powering the Wu Dao models.

The first large-scale generative model for text-to-image synthesis in the general domain for Chinese. CogView is a 4-billion-parameter Transformer usi…

China's first large-scale model family. Includes Wen Yuan (2.6B language), Wen Lan (1B multimodal), Wen Hui (11.3B generative), and Wen Su (protein fo…