TAAFT
Free mode
100% free
Freemium
Free Trial
Prompts Deals
Companies
Beijing Academy of Artificial Intelligence (BAAI)
Models
Orca (BAAI)

Orca (BAAI)

By Beijing Academy of Artificial Intelligence (BAAI)
Orca reframes world modeling around Next State Prediction instead of next token, next frame, or next action objectives. It learns a shared world latent space from multimodal signals, mainly continuous video and language, through two complementary learning modes: unconscious learning captures dense natural dynamics from video without labels, while conscious learning organizes observations into causal structure using language, events, and questions. Once learned, the frozen world latent is decoded through lightweight readouts for text generation, image based future state prediction, and robot action generation. Orca comes in 0.8B and 4B backbone sizes and has been evaluated on video question answering benchmarks, the PRICE physical prediction benchmark, and real dual arm robot manipulation tasks, showing strong performance relative to similarly sized vision language models and world models.
New Multimodal
Released: June 29, 2026

Overview

Orca is a world foundation model that learns a unified latent world state from multimodal signals to predict future states, rather than predicting next tokens, frames, or actions directly. It uses unconscious learning from continuous video and conscious learning from language described events to model physically consistent state transitions, which are read out through language, vision, and action interfaces.

🎥Video analysis🔍Research🔍Image analysis🤖Robotics

About Beijing Academy of Artificial Intelligence (BAAI)

Beijing Academy of Artificial Intelligence (BAAI), also known as Zhiyuan Institute, is a Chinese non-profit AI research laboratory established in November 2018. BAAI conducts fundamental AI research, develops open-source models (including WuDao, BGE embeddings, Emu, and RoboBrain), and fosters collaboration between academia and industry. Known for creating the BGE (BAAI General Embedding) series of embedding models used in RAG systems worldwide.

Industry: Research
Company Size: 260
Location: Beijing, Beijing, CN
Website: baai.ac.cn
View Company Profile

Related Models

  • Marlin 2B
    Marlin-2B is NemoStation’s open-source 2B video-language model for dense video captioning and natural-language temporal grounding.
    🎥Video analysis🔍Video search📝Video captions
    NewMultimodal
    Released 1mo ago
  • Protenix
    Protenix is ByteDance’s open-source biomolecular structure prediction model family for high-accuracy protein and complex structure modeling. It is designed as an accessible foundation for computational biology research, supports protein, RNA MSA, and template features in its latest release, and is positioned as a strong open alternative for structure prediction and downstream biology workflows.
    🧬Protein design🔬Scientific research🔬Biology research assistance🔬Protein engineering analysis
    Multimodal
    Released 5mo ago
  • Aurora
    Aurora is Creatify’s diffusion-based multimodal avatar video model for turning a single image plus audio or text into ultra-realistic speaking or singing videos. It is built for studio-grade ad creation and reactive digital avatars, with strong lip sync, emotional expression, hand gestures, and long-form character consistency from just one reference photo.
    Multimodal
    Released 1y ago
Last updated: July 9, 2026
0 AIs selected
Clear selection
#
Name
Task