iFlytek launches Spark X2.6: a 293B-A30B MoE with 1M context that it says was trained entirely on Chinese chips, priced at ¥0.9/¥3.6 per million tokens; only the 4B sibling has open weights

Why it matters

【Key points】

At its 1024 Developer Festival in Beijing on Oct 22, iFlytek launched Spark X2.6, a 293B-total / ~30B-active MoE model with a 1M-token context, aimed at coding and agent tasks. iFlytek says it was trained entirely on Chinese-made compute. Chairman Liu Qingfeng claimed Spark is the only mainstream Chinese model not trained on NVIDIA GPUs.

Launch pricing (40% off) is ¥0.9 per million input tokens (¥0.08 cached) and ¥3.6 per million output tokens, about $0.13/$0.51. The flagship's weights are closed. Its sibling Spark-X2.6-4B is open under Apache 2.0, with 1M context and three reasoning levels. iFlytek gave no benchmark scores for the flagship.

【Verifiable】

https://xinghuo.xfyun.cn/sparkapi

https://www.ithome.com/1/016/403.htm

https://1024.iflytek.com/

https://huggingface.co/XHToken/Spark-X2.6-4B-Base

【Our read】

1. iFlytek has trained Spark on Huawei Ascend clusters since 2023. X2.6 keeps the 293B MoE size but adds 1M context and shifts toward coding and agents.

2. Its promo price is close to MiniMax M3.1 Flash ($0.15/$0.60) and below DeepSeek V4.1 Flash ($0.30/$1.20). Unlike those two, it has no published benchmarks and no open flagship weights. DeepSeek reportedly plans to start Ascend pretraining by year-end, while ByteDance, Alibaba and Tencent are buying NVIDIA's export-compliant RTX PRO 5500.

3. A training stack free of export controls appeals to government buyers. For developers, independent evaluations will decide whether it matters. The "only all-domestic" claim and the training scale are unverified.