JT-35B-Flash

China Mobile's "Jiutian 35B" general model as a served API: a 35B-parameter, non-reasoning, text-only model with a 256K-token context, released May 14, 2026 and positioned by Artificial Analysis at launch as "a proprietary 35B non-reasoning model with relatively high token efficiency and competitive intelligence for its size" and "a significant upgrade" over China Mobile's earlier models. It scored 36 on AA's Intelligence Index v4.1 at release; 29 on v4.1.1, and on the current v4.3 it stands at 19, well above the ~8 median for non-reasoning models in its price band. Two months later the larger JT-4.1 Flash (236B-A21B, 40) took over the top of the Flash line.

The 35B base is documented in China Mobile's JT-Safe paper (arXiv 2510.17918): JT-35B-Base was pretrained in-house on 6.2T tokens, then continue-pretrained on 1.5T tokens of "Data with World Context" to improve safety and factuality, and the model passed China's dual generative-AI filing and A-level security certification. Ahead of its formal unveiling at the 9th Digital China Construction Summit, China Mobile announced deep inference adaptation of the 35B model on Huawei Ascend 910 (April 27, 2026) and Moore Threads' MTT S5000 GPUs, with Biren also reporting pre-adaptation — a domestic-silicon deployment story typical of the state-owned telecom's AI line. Weights are not public; dense/MoE layout, training hardware and API pricing are undisclosed (AA lists it at $0 pending a published price).