Ant Group open-sources 560-billion-parameter Ling-3.1-flash AI model

Summary

Ant Group released the weights for Ling-3.1-flash, a 560-billion parameter MoE model. The model utilizes approximately 25 billion active parameters per token and offers a 1-million token context window. Focused on coding, agentic workflows, and long-document processing, the model can be deployed via SGLang and vLLM.

Why it matters

A major Chinese player releasing a 560-billion parameter open-weights model pushes the open-source frontier even further. Mid-sized teams gain a credible alternative to closed APIs for agentic and coding workloads. The mixed-precision technique, which halves GPU requirements (from 8 to 4 Hopper GPUs), effectively lowers the barrier for self-hosting. This move also signals Ant Group's strategic shift from payments to foundational AI infrastructure, intensifying competition with Alibaba, ByteDance, and DeepSeek.

Highlights

• Total 560B / Active 25B (MoE), 1M token context, SGLang and vLLM compatible
• Mixed precision reduces weight memory from 534 GB to 292 GB; drops GPU requirement from 8 to 4 Hopper GPUs
• Focused on professional tasks involving coding, agent tool use, and long documents

Critical take

The article lacks independent benchmarks, perplexity scores, or direct comparisons with models like Qwen3 and Llama 4. Consequently, the 'flash' performance claims remain entirely unverified. Additionally, it fails to address critical details for any team planning deployment, such as licensing, fine-tuning restrictions, and commercial usage terms.