Ant Group's Ling 3.1 Flash is out: ~560B MoE, weights promised 'soon'

Ant Group's Ling team announced Ling-3.1-flash today, and the model already answers on a public API: inclusionai/ling-3.1-flash is listed on the Vercel AI Gateway and on nano-gpt.

WHAT ANT SAID

From Ant Ling's own post at 16:33 UTC: roughly 560B total parameters with about 25B active per token, a context window of up to 1M tokens, and "We plan to open-source the model soon." Its own benchmark figures: 1,673 Elo on GDPVal-AA v2.1, 75.16 on FrontierSWE and 65.35 on HealthBench Professional. Those are Ant's numbers, not a third-party result.

WHAT THE CATALOGS SAY

models.dev lists Ling 3.1 Flash as a reasoning model released September 29, with a 262,144-token context and 32,768 max output on both listings, well short of the 1M Ant describes. The Vercel gateway carries it at no charge for now; nano-gpt charges $0.075 in and $0.22 out per million tokens. Ant has not published a price of its own.

WHAT WE DO NOT KNOW

When the weights land. As of this afternoon there is no Ling-3.1 repository under inclusionAI on Hugging Face, so posts calling this an open-source release are early. It is not on OpenRouter yet either, where Ling 3.0 Flash first appeared in July.

ON OUR END

The radar saw ling-3.1-flash and filed it as an early variant, because the entry was named "Ling 3.1" and "flash" reads as a variant marker. Ant ships this line as Flash. The entry is renamed and a test pins it: the same miss we fixed for Gemini 3.8 Flash on September 2, in a family that fix did not reach.