Ling-3.1-flash was officially open-sourced by Ant Group's Ling team on October 15, 2026, with weights listed on Hugging Face and ModelScope. The model had launched on September 30 with a trial period that ended on October 14; what lands today is the full weights plus the complete one-million-token context, which is the step that actually lets teams wire the model into their own systems.
Start with the deployment bar: 560B in size, 25B active
Ling-3.1-flash continues the hybrid linear architecture of Ling-3.0-flash, with roughly 560B total parameters, about 25B active per token and a context ceiling of 1M tokens. The official model card's reference deployment is eight H200 GPUs or four GB300 GPUs, with serving paths for SGLang and vLLM. This is plainly not built for a personal computer, but its activation ratio cuts the compute of each inference compared with a dense model of similar size. It pulls in the opposite direction from MiniCPM-V 4.7 squeezing a multimodal model into phone memory: one shrinks the model until the edge can run it, the other makes a large model affordable to run and use inside a data center.
Read the benchmarks in pieces: strength in office work and delivery, not just a total score
In Artificial Analysis public testing, Ling-3.1-flash posts an intelligence index of 41, output speed of about 210 tokens per second and a blended price near $0.19 per million tokens. Its scenario scores say more about positioning: 1673 on the GDPVal-AA V2.1 office-work benchmark and 65.35 on HealthBench Professional for medical scenes, with corresponding results on software-engineering and terminal-operation benchmarks. In the publisher's hands-on demos, it turned scattered earnings material into editable tables and an interactive page, and read legacy code, wrote new features and kept fixing them against test results. The point of this release is less the parameter count than the chance for teams to test, on their own task sets, how often it actually finishes real work handed to it in fuzzy business language.
Before wiring the weights into production, three checks are worth doing first. First, run a small-sample comparison on your own office and engineering tasks, watching whether it can recover mid-task after a mistake rather than judging single-turn answers. Second, measure cost at your real share of full 1M-context calls, because bills and latency move noticeably when context is maxed out. Third, in high-risk areas such as medicine, keep it to auxiliary steps like organising histories and question lists, with formal conclusions still going through professional review. Only after those three steps does it make sense to judge whether it should replace the model in an existing workflow.
