Ling 3.1 Flash 1019
InclusionAI · 2026-10-27
InclusionAI's further post-trained update of Ling 3.1 Flash, published as the separate MIT-licensed FP8 checkpoint Ling-3.1-flash-1019 on Hugging Face on October 27, 2026 and also served through InclusionAI's first-party API; its card reports improved reasoning and agentic performance, with thinking toggled by enable_thinking. Its config.json is byte-identical to Ling 3.1 Flash's (hybrid-linear MoE, 63 KDA and 9 gated MLA layers, 512 routed experts with 8 active), and the card gives the same rounded 560B total and 25B activated parameters, with no exact unquantized count. The config sets 262,144 native positions, while the card advertises a 1M-token context window as Ling 3.1 Flash's does.