Ant Open Source

Ant Group open-sources a family where both the understanding backbone (LLaDA2.0-mini) and the 6B DiT generator are diffusion models trained in one framework. Over 90% of roughly 220M generation samples used image-only supervision, with text pairs added later for alignment. A Turbo variant runs 2-4 s

Ant's new flagship instant model activates 63B of a trillion parameters, trained on a 29T token corpus with a 1M token context, with reported upgrades to architecture, token efficiency and preference alignment. Weights are on Hugging Face and ModelScope.

LLaDA2.0 pushes discrete diffusion language models to 100B parameters with open weights for both 16B and 100B MoE variants. Parallel decoding hits 535 tokens/s in SGLang, and the team details the Warmup-Stable-Decay strategy that converts autoregressive models into diffusion ones.

The boba release of Ant's asynchronous RL system reports a 7B model at 61.9 on AIME24 and 48.3 on AIME25, and a 32B matching QwQ on AIME24 from only 200 samples, with full SGLang support. Training data and weights are released alongside.

Ant open-sources AEnvironment, the environment layer under its agentic RL training. It encapsulates agents as environments so multi-agent setups compose, supports large-scale parallel sampling for trillion-parameter runs, synthesizes training trajectories, and ships MCP and benchmark support.

Region-to-Image distillation trains a vision language model to resolve fine detail internally rather than calling a crop-and-zoom tool. ZwZ-8B reports state-of-the-art fine-grained perception with zero tool calls, and the release adds ZoomBench with a dual-view zooming gap metric.

Ant's TuGraph team introduced Chat2Graph, a graph-native agentic system pairing a multi-agent runtime with a graph database so users can explore and build graphs conversationally instead of writing queries. Source is on GitHub.

DR-Venus is a 4B deep research agent trained entirely on open data through strict-cleaned SFT then RL with turn-level IGPO optimization and dense credit assignment across trajectories over 200 turns. Training code, GGUF checkpoints for edge deployment and a paper are all released.

Ant open-sources AState, the state data layer from its Ling-1T RL training stack. Automatic weight resharding with zero-redundancy in-place updates drops trillion-parameter train-to-rollout synchronization from minutes to about six seconds, with NUMA-aware placement for thousand-card jobs.

AReaL v1.0 puts a transparent proxy in front of an agent's model endpoint, capturing agent and LLM exchanges as RL training data without modifying the agent. It ships the Archon engine with PyTorch-native 5D parallelism and reports 73.0% pass@1 on tau2Bench Airline, 98.3% on Telecom.

LLaDA2.0-Uni handles multimodal understanding and generation as one block-wise mask prediction task on an MoE diffusion backbone. It reasons in text before generating images, encodes images as purely semantic discrete tokens, decodes in eight steps, and scores 0.78 on WISE-Bench.

Ling-2.6-1T is Ant's new trillion-parameter flagship for agentic use, positioned around intelligence per token, engineering task performance and production readiness. Weights are on Hugging Face and ModelScope.

Ant and InclusionAI published a quarterly map of open agentic AI: 50+ projects sorted into coding agents, personal assistants and orchestration frameworks, alongside contribution data from 21K+ developers showing an extreme power law and under 10% of activity coming from big tech.

Ant open-sources AMem, an NCCL plugin that makes communicator memory offloadable through ncclPause and ncclResume. It frees more than 10GB per Hopper GPU and cuts the train-to-inference phase switch from minutes to under a second by preserving the group and restoring only metadata.

Ant open-sourced Ling-2.6-flash, a model aimed at agent workloads with a hybrid linear attention architecture and optimization for spending fewer tokens per task. Weights are available on Hugging Face and ModelScope.

Ant's updated trillion-parameter reasoner moves to a 1:7 mix of multi-head latent attention and lightning linear attention, targeting faster reasoning and cheaper multi-step tool collaboration. Weights are published on Hugging Face and ModelScope.

Awex is an open weight synchronization framework for RL pipelines, pushing updated parameters from training to inference and reporting 1T-parameter syncs across 1000-GPU clusters in under six seconds. It transfers only needed shards, updates in place, and supports NCCL, RDMA and shared memory.

Ant open-sourced AntAngelMed, a 100B-parameter MoE model built on the Ling architecture and trained on large-scale real-world medical data. It claims the top spot among open models on HealthBench, putting a domain baseline at that scale in reach of self-hosting.

AReaL v0.5.0, the fully asynchronous RL engine behind Ring-1T post-training, is out. It decouples agentic RL behind an OpenAI API proxy so agents train without bespoke rollout plumbing, and swaps SPMD for a single controller to kill long-tail latency and data imbalance.

Ant's AI Security Lab and Tsinghua University released ClawAegis, a lightweight guard layer for autonomous agents that installs as a native plugin and defends across the agent lifecycle, targeting malicious instruction injection and memory poisoning.

LLaDA2.1-flash, the 100B token-editing diffusion model, is now served on ZenMux via API and chat at a quoted 892 tokens per second. Hosted access removes the GPU provisioning step for teams wanting to test diffusion generation against their current models.

LLaDA2.1 pairs block diffusion with an error-correcting editable engine, so the model revises drafted tokens instead of committing to them. Released in 16B and 100B sizes with SGLang support, reporting 892 tokens per second on HumanEval+ and 801 on BigCodeBench.