Ant Open Source

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 post-trained Ling-3.0-tiny on a single DGX Spark using AReno and an agentic RL tic-tac-toe environment driven by tool calls and reward feedback. Mean reward moved from roughly -0.5 to 0.4, response length settled near 850 tokens, and behavior stabilized, sketching a local train-and-adapt loop fo

Ant and the SGLang community ran a full DeepSeek-V4-Pro serving project through Infer-forge, an agentic harness combining Harness, Loop and Graph. It split the work into 38 verified pieces, held a 28-hour median task lifetime traceable, lifted peak concurrent tasks from 2 to 9, and caught silent ker

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.

Ant Group's open-source team released FP8, FP4, and INT4 quantized builds of Ling-3.0-flash-Fin, a finance-tuned model, on Hugging Face and ModelScope, so teams can trade accuracy for memory and speed based on their deployment target.

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.

Ant and the SGLang team detail how they serve DeepSeek-V4-Pro on H20 with scenario-specific profiles: Humming MXFP4AFP8 weights, online C128 KV compression, DSpark speculation across pipeline stages. Reports 271 tok/s at batch 1 and 1M-token prefill in 43.7s, with benchmark settings.

LLaDA2.2-flash is an MoE diffusion language model aimed at agent workloads, using Levenshtein style edits (deletion, insertion, substitution) to self correct mid decode rather than reverting to left to right generation. Reported: 703.82 TPS on BFCL-V4 function calling and 1.64x BF16 throughput.