Ant Group

Unified discrete diffusion language model that natively combines multimodal understanding and generation in a single architecture. 16B MoE with only ~…

Latest Ling iteration featuring Lightning Linear Attention, supporting a 1 million token context window and optimized for agentic interaction.

Perception benchmark for fine-grained multimodal understanding (Zooming-without-Zooming).

First industry model to unify speech, audio, and music within a single architecture.

Scales diffusion language models to 100B by converting pretrained AR models. LLaDA 2.1 introduces token-to-token editing for real self-correction, ach…

Trillion-parameter MoE flagship with ~50B active parameters. The Ling 2.0 technical report details reasoning-oriented MoE foundation scaling from tens…

Ant Group's reasoning-oriented (thinking) model family, built on the Ling backbone with reinforcement learning. The world's first open-source trillion…

MoE scaling laws paper exploring efficient mixture-of-experts language models.

Massive dataset for training financial reasoning agents.

Evaluation set for human-centric AI and embodied intelligence.

Fully asynchronous reinforcement learning system for training large reasoning and agentic models. Achieves up to 2.77x training speedup compared to sy…

Ant Group's native multimodal model built on the Ling backbone. Handles vision, speech, audio, and music.

Multi-agent framework for building, evaluating, and training general multi-agent assistants. Top-ranked open-source framework on the GAIA benchmark. S…

Ant Group's general-purpose (non-thinking) LLM family, introduced with "Every FLOP Counts" — a technical report on training massive MoE models without…

Experimental diffusion language model from Inclusion AI. Explores an alternative to autoregressive generation. LLaDA-MoE extends the approach with Mix…

Papers accepted at CVPR and AAAI focused on high-fidelity human animation and multimodal generation.

Comprehensive multi-agent framework designed for industrial-scale agentic AI.

Fundamental architectural research allowing the Ling series to process massive context windows with minimal memory overhead.

High-performance vector indexing library for similarity searches in RAG (Retrieval-Augmented Generation).

First specialized model fine-tuned for the finance industry, trained on over 1 trillion tokens.

Specialized LLM series for the full software development lifecycle (coding, testing, and DevOps).

Benchmark specifically designed to evaluate LLMs in the Chinese financial domain.

Tools for Knowledge Graph (SPG) and Large-Scale Graph Learning.

Library for interacting with databases using private large-model technology.

Automated distributed deep learning system that optimizes resource allocation for training.

High-performance graph database used for risk control and fraud detection.

Unified framework for privacy-preserving machine learning (MPC, TEE, HE).

Globally leading scalable distributed database. Core infrastructure for Ant's AI deployments.

Secure container technology co-founded by Ant employees, used for isolating AI workloads.