Ant Group launches Ling-3.1-flash with 560 billion parameters

Summary

Ant Group's InclusionAI team introduced Ling-3.1-flash, a 560-billion parameter MoE model designed for agent, search, and office workloads. The model activates approximately 25 billion parameters per token. While the design targets a 1-million-token context window, this limit is capped at 256,000 tokens during the two-week free trial period.

Why it matters

A 560-billion parameter model from a major Chinese fintech player demonstrates that the frontier model race is no longer limited to a handful of AI labs. The stated intent to release the weights as open source after the trial, if fulfilled, would add a significant open-weight option for agent and long-context workloads. The 1-million-token context target also pressures competitors to scale up their support for very long windows.

Highlights

• 560 billion total / ~25 billion active parameters (MoE architecture)
• 1-million-token context window in design; however, the trial is limited to 256,000 tokens
• Positioned for agents, search, office software, and specialized apps; open-source release promised after the trial

Critical take

The article is a bare two-paragraph news brief. It lacks benchmarks, head-to-head comparisons, training data details, or evaluation specifics. As a result, the '560B' headline figure says very little about actual capabilities. The open-source promise is softened with the phrase 'planned.' Additionally, the 256,000-token trial limit means early users can only test a quarter of the advertised context. Under these conditions, it is still too early for a real-world assessment.