Ling-3.1-flash

Ling-3.1-flash is the next-generation efficiency-focused Mixture-of-Experts model from Ant Group's inclusionAI lab, unveiled in late September 2026 (2026-09-29/30) as the successor to Ling-3.0-flash. It scales to 560B total parameters activating about 25B per token — more than quadrupling the 124B total / 5.1B active of Ling-3.0-flash — and is a reasoning model with an explicit thinking mode, built for agent tasks, search, office software, and specialist applications. Text input and output; a 262,144-token (256K) context in the launch/trial hosting, with 1M cited as the target once the full window is enabled, and up to 32,768 output tokens. It launched as a two-week free trial via Vercel's AI Gateway (hosted by Novita), with inclusionAI stating open weights will follow after the trial — so as of launch the weights, license, and any post-promo per-token price are announced but not yet posted (not self-hostable today). inclusionAI-tracked benchmarks include an aggregate agentic score of 81.0 across five agent benchmarks, Terminal-Bench 4 40.4%, SWE-Atlas 55.9%, and HealthBench Professional 65.3%; no single public overall score was available at launch. Figures are vendor/third-party-reported.