The news isn't just in the benchmarks, but in the confirmation of a precise architectural trajectory. inclusionAI has published Ling-3.1-flash-1019, an updated version of Ling-3.1-flash, keeping 560 billion total parameters and 25 billion active parameters per token, but pushing the context window to one million tokens. The improvements focus on agentic tasks: AA-Briefcase moves from about 1400 to 1524 Elo, AutomationBench-AA from 62% to 68%. Scientific benchmarks also rise: CritPt from 18% to 26% and SciCode from 54% to 57%. The Artificial Analysis Intelligence Index v4.3.2 climbs from 41 to 44.

Memory before compute

For those evaluating self-hosted deployment, these numbers hide a tension. On one hand, 25 billion active parameters per token is a value compatible with models that run on high-end single GPUs; the computational cost per token remains relatively contained. On the other, the full 560 billion parameter checkpoint cannot be ignored: every expert, even if not active, must be loaded into VRAM or retrieved quickly. This shifts the bottleneck from compute to memory capacity and bandwidth, the opposite of what happens with smaller dense models.

The one-million-token context window amplifies the problem. Such a wide window can turn the key-value cache into a critical memory object, especially if you want to keep latency low without relying on offloading to slow storage. Serving pipelines must therefore manage two planes: expert loading and conversation state. It is no coincidence that the base checkpoint, Ling-3.1-flash-base, is released for research: the community can experiment with quantization and expert scheduling strategies without starting from scratch.

The jump on AA-Briefcase and AutomationBench-AA indicates that post-training optimization didn't just improve answers to complex questions, but made the model more reliable in multi-step workflows. This is what matters to those doing local deployment: serving tokens fast isn't enough; you need a stack capable of maintaining coherence over long tasks without degrading latency.

The release of the base checkpoint signals that inclusionAI is not just aiming to beat benchmarks: it wants to shift the discussion on sparse architectures as an open research ground. For those building local stacks, the question isn't whether to adopt this model, but whether their memory and serving infrastructure is ready for a generation in which total parameter count matters less than the number of active parameters.