Overview
At the 10th iFlytek Global 1024 Developer Festival, Chairman Liu Qingfeng released iFlytek Spark X2.6, trained entirely on domestic compute. The model uses a 293B-total / 30B-active MoE architecture with context extended to 1 million tokens, marking the first time code and agent capabilities break through on a fully domestic compute base.
The bigger story is the “edge-cloud integration” engineering paradigm: the cloud side is a 293B-total / 30B-active MoE model trained and inferred entirely on domestic accelerators; the edge side is an open-source 4B small model running on phones, PCs and even learning devices; a routing layer decides per task whether to run on device or in the cloud. iFlytek disclosed that the share of tasks completable independently on device rose to 41.9%.
Background & Interpretation
1. Technical Background and Evolution
Large-model development is moving from centralized cloud inference to edge-cloud collaboration. Early speech input required uploading to a server for recognition; today on-device dictation is routine. Models are retracing that path, only at far greater scale. A full agent task can span dozens of calls and consume hundreds of thousands of tokens; relying entirely on the cloud imposes steep compute cost and data-compliance pressure.
Domestic accelerators have long been constrained by memory and bandwidth, making trillion-parameter training harder. iFlytek’s engineering answer is two in-house paths: “long-horizon dynamic parallel” training lifting ultra-long-context training efficiency about 4x; and inference-side “adaptive compute-storage collaboration plus four-layer KV Cache plus distributed inference” lifting inference efficiency 4.5x.
2. Core Drivers and Underlying Mechanisms
Edge-cloud integration is fundamentally a rebuild of “compute economics”: frequent, low-value calls sink to the free on-device tier, while the cloud handles only complex tasks, rewriting the token-consumption structure. What lets the router judge where a task goes? Real-time assessment of task length, privacy sensitivity and compute demand.
For domestic chips, this means moving from “usable” to “practical”—not racing top foreign chips on peak FLOPS, but running full-stack trillion-parameter training and inference on constrained hardware through system-level collaboration (compute-storage synergy, KV Cache compression, distributed inference).
3. Market Structure and Industry Chain Impact
Spark X2.6, together with contemporaneous Ascend training practices at Shenzhen’s Hetao academy, shows domestic AI chips can now support full-stack training and inference of hundred-billion-parameter models. This reduces dependence on overseas compute and ecosystems, and provides a local model base for edge forms like all-in-one machines, AI phones and AI PCs.
On the developer front, iFlytek’s AI developer contest drew 52,000 teams from 351 cities in 15 countries, 43% from non-technical backgrounds, signalling that model deployment is shifting from “geek-exclusive” to “build-for-everyone.”
Implications & Outlook
Near term, edge-cloud integration will reshape enterprise model-selection logic: whoever can safely, stably and controllably take over real business workflows holds the growth ticket. Medium term, as vendors like DeepSeek keep compressing KV Cache and HBM demand, per-token cost will show generational gaps, making “cheap enough to dare use” the core competitive edge of domestic models.
Key variables: the scaled delivery capability of domestic compute clusters, the collaboration standard between edge small models and cloud large models, and the penetration speed of industry models into verticals like ERP, law and education.
Takeaways for Industry Participants
- Enterprise IT should evaluate an “edge-cloud split” architecture, keeping frequent low-sensitivity tasks local to control cost and compliance risk.
- Chip and hardware vendors can differentiate on compute-storage synergy and KV Cache compression rather than chase peak FLOPS.
- Developers should watch low-code / no-code agent-building tools to package domain know-how into callable capabilities.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.