Ling 3.1 Flash

Ling 3.1 Flash is a large mixture-of-experts model released by Ant Group's InclusionAI, built around a 560 billion total parameter design with roughly 25 billion parameters activated per token. That sparsity profile lets it deliver substantial capacity while keeping per-token compute closer to a mid-sized dense model, which is a common pattern for agent and productivity workloads where throughput matters. InclusionAI has positioned the release for agent tasks, search, office software, and other specialist applications, suggesting a focus on tool-using, instruction-following behavior rather than open-ended creative generation.

The model is designed to handle very long contexts, with a stated target window of up to one million tokens for the underlying architecture. At launch, a two-week free trial exposed it through a 256,000-token cap, and InclusionAI has indicated it plans to unlock the full one-million-token window and publish the weights as open source once that trial concludes. For practitioners, this means Ling 3.1 Flash is best understood today as a frontier-scale MoE being staged for broad availability, with its long-context and open-weights potential as the main forward-looking advantages once the broader release lands.