--- title: "iFlytek bets on fully domestic compute with Spark X2.6 and an end-cloud split" date: 2026-10-23 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-iflytek-spark-x26-domestic language: en --- # iFlytek bets on fully domestic compute with Spark X2.6 and an end-cloud split > iFlytek's Spark X2.6 runs entirely on homegrown chips and routes work between device and cloud to cut cost and leakage. A Chinese AI lab just made a claim that sounds boring until you sit with it: it says its newest large model (大模型) was trained from scratch on domestic chips, not Nvidia GPUs. For most of the past three years, that sentence would have been a punchline. At the 2026 iFlytek 1024 Global Developer Festival in Beijing, chairman Liu Qingfeng (刘庆峰) put it at the center of the launch anyway — and dared the room to test it. The model is Spark X2.6 (星火 X2.6). The bet behind it is not raw capability. It is that the next competitive moat in Chinese AI is not "who scores highest on a benchmark" but "who can ship a model that runs on hardware you actually control, keeps your data on your own devices, and still does real work." ## What Spark X2.6 actually is Spark X2.6 is a mixture-of-experts (混合专家, MoE) model rated at 293B-A30B — 293 billion total parameters, 30 billion active per token. It supports a one-million-token context (上下文) window and was, per iFlytek, trained and run on a homegrown 10,000-card inference cluster. The more interesting part is the companion release: **Spark X2.6-4B**, a 4-billion-parameter on-device model that iFlytek open-sourced. Where the 293B flagship lives in the cloud, the 4B sibling is meant to run on phones, PCs, tablets, AI glasses, car cockpits, and robots. - Cloud model: deep reasoning, long-horizon agent (智能体) tasks, multimodal (多模态) understanding. - Device model: search, Q&A, document handling, scheduling, local spreadsheet work. - The two are stitched together by a routing layer that decides, per request, where the work should happen. ## The end-cloud (端云一体) architecture is the real story iFlytek is framing the launch around a single idea: stop sending everything to the cloud. Its new end-cloud collaborative architecture uses an on-device "precise router" that classifies each task as pure-on-device, pure-cloud, or a hybrid, based on difficulty, device capability, and data-sensitivity needs. The company says the router hits **96.1% accuracy**, and that sensitive fields are desensitized locally before anything leaves the device — with the original file never uploaded. Liu's own back-of-envelope split: roughly 60% of scenarios could run on-device, 30% via end-cloud coordination, and 10% in the cloud. That framing answers two complaints enterprises repeat constantly: - **Fear of leakage** — keep regulated or proprietary data local. - **Fear of cost** — stop burning tokens (and therefore money) on tasks a phone can do. iFlytek also claims concrete efficiency gains from working inside domestic hardware limits: long-context training efficiency up about 4x, inference efficiency up 4.5x, and KVCache hit rate lifted from 63% to 95.7%. These are company-reported engineering figures, not independently audited, and should be read as iFlytek's own account of its work. ## Where it is already deployed iFlytek pushed X2.6 into the fields it knows best — education, healthcare, law, and government services — rather than selling it as a generic chatbot. - **Legal**: its "Xiaofa" (晓法) assistant, built on the legal large model plus X2.6's end-cloud design, covers case analysis, evidence review, litigation strategy, and contract review. iFlytek says it lifted overall case-handling efficiency by 45% in demos. - **On-device**: the X2.6-4B model improves coding, agent work, long-text handling, and multilingual performance over the prior X2.5-4B, with an agent task coverage the company measures at 41.9%. - **Ecosystem**: the open-source 4B model had drawn 2.56 million cumulative downloads since the X2.5-4B release in September, a figure iFlytek cites as placing it among the top four on-device models globally by download volume. ## Pricing and the cost angle iFlytek listed API rates for the cloud model at 0.9 yuan input, 0.08 yuan cached input, and 3.6 yuan output per million tokens — roughly **US$0.13 / HK$0.98**, **US$0.01 / HK$0.09**, and **US$0.51 / HK$3.9** respectively at ~7.1 RMB to the dollar. Cached input is heavily discounted, signaling a strategy built around long, repeated contexts rather than one-shot prompts. ## Trend reading According to Li An, chief scientist at BrainNet (脑机网), "the labs that win the next two years will not be the ones with the biggest parameter counts, but the ones that can prove their stack runs end-to-end on hardware their customers are allowed to buy and keep." ## Honest limitations - The "trained entirely on domestic computing" claim is iFlytek's own statement. No independent benchmark currently verifies that the training run matched what the same model would achieve on equivalent Nvidia-class hardware. - Efficiency numbers (4x training, 4.5x inference, 95.7% cache hit) are company-reported and lack third-party reproduction. - Liu Qingfeng himself conceded that domestic compute still trails the global frontier by about two generations — so "autonomous and controllable" (自主可控) is a direction, not a finished parity claim. - X2.6-4B is open-weight; the 293B flagship is delivered as a hosted, API-facing service, not published weights. Buyers should not assume the flagship is self-hostable. - The 45% legal-efficiency figure comes from iFlytek demos, not an external audit. ## What readers can do now - If you handle regulated data in China, test the X2.6-4B open model locally first — it is the part you can actually audit and keep on-premise. - For cloud pilots, benchmark the cached-input price against your real long-context workload before assuming the discount applies to you. - Treat the "fully domestic compute" line as a procurement and compliance feature to verify, not a performance guarantee to quote. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-iflytek-spark-x26-domestic Free to quote with attribution and a link to the original.