JT-4.1 Flash

China Mobile's Jiutian LLM line reaches frontier-adjacent territory: JT-4.1 Flash is a 236B-total / 21B-active MoE (non-reasoning) with a 256K-token context. AA Intelligence Index v4.1: 39, ranking near the top of its non-reasoning class — a notable result for a model from a state-owned telecom's in-house AI brand rather than a dedicated frontier lab. Released 9 July 2026.

Weights are not publicly available; the 21B active-parameter figure follows China Mobile's "236B A21B" model-naming convention. The 236B/21B MoE config matches DeepSeek-V2's published architecture, but the Jiutian team's documented pattern is adopting open architectures while pretraining from scratch (JT-Coder-8B-Base is stated "from-scratch" with a Qwen-compatible design, and JT-Safe/arXiv 2510.17918 documents the sibling JT-35B-Base's full in-house pretraining pipeline on domestic clusters) — no source claims weight-level derivation.