Ling-3.1-flash-1019

A post-training upgrade of Ling-3.1-flash, with the same hybrid-linear MoE: 560B total and 25B active parameters, 63 KDA and 9 gated MLA layers, and 512 routed experts plus one shared, 8 active per token. It first appeared on Artificial Analysis on October 22, 2026 as an inclusionAI API model (release slug ling-3-1-flash-rc5), and the weights followed on HuggingFace on October 27 under the MIT license. The card advertises a 1M-token context, as for the original, where 262,144 is native and 1M comes from a YaRN override. inclusionAI attributes the gains to "further post-training optimizations". The card's figures are all Artificial Analysis evals: AA-Briefcase rises from about 1400 to 1524 Elo, AutomationBench-AA from 62% to 68%, CritPt from 18% to 26% and SciCode from 54% to 57%.

Artificial Analysis scores it 30 on Intelligence Index v5.0.1 (29.6 unrounded; 29 on v5.0 before AA regraded CritPt v2 in November), three points behind DeepSeek-V4.1-Flash-1017 (33) and level with MiMo-V2.6-Pro (30, 29.9 unrounded) among open-weights models. On v4.3.2 it scored 44, three points above the original release. It is strongest on AutomationBench-AA (67.5%) and the CritPt v2 hard subset (37.5%, the best of any open-weights model AA has scored) and weakest on Terminal-Bench-Science (4.3%). It is cheap to run, at $3.19 per index task on $0.30 / $0.90 per million tokens.

Two days before the 1019 weights, on October 25, inclusionAI also released Ling-3.1-flash-base, the BF16 pre-trained checkpoint behind both releases (trained in FP8, then merged with Warmup-Stable and Merge). Its card compares it with DeepSeek-V4-Pro-Base and MiMo-V2.5-Pro-Base on 47 benchmarks.