On 22 October 2026, at the company's 1024 developer festival, iFlytek chairman Liu Qingfeng presented Spark X2.6 — a model trained entirely on Chinese compute infrastructure. He put the gap to NVIDIA at "nearly two generations", mostly in memory and bandwidth. The company's answer is not an attempt to close that gap but to move part of the work onto a model running locally on the device.
Key takeaways
- Spark X2.6 is an edge-cloud pairing: one large cloud model plus a 4B model running on the device
- Routing accuracy between device and cloud: 96.1% per company figures
- The 4B model raised the completion rate on complex agent tasks from around 12% to 41.9%, with 60% as the next target
- Long-context training was sped up roughly 4×, and the inference side 4.5×
- Liu: "Right now only Huawei and us are doing this as a main effort" — on training on Chinese accelerators
Let us picture a situation
A procurement team has a contract to review. The document holds prices, terms and counterparty details that must not leave the building. A model could help, but to read the contract it would have to send it to the cloud.
The demonstration shown at the festival does it differently. Sensitive passages are anonymised locally first, the user confirms, and only then does some data travel onward. The original file never leaves the device.
But here is the problem: a model small enough for a laptop is orders of magnitude smaller than a cloud model — and until recently it could not handle multi-step work.
That is the metric iFlytek is trying to lift.
What exactly was announced?
Spark X2.6 is not one model but a pair. A large model runs in the cloud, and a 4-billion-parameter model on the device. A routing layer decides where a task runs, on two criteria: task difficulty and data sensitivity.
A task can stay entirely local, be split between device and cloud, or go wholly to the cloud. The company puts the accuracy of that routing at 96.1%.
| Metric | Value | Reference point |
|---|---|---|
| Complex agent-task completion (4B model) | 41.9% | approx. 12% previously in the industry, 60% targeted |
| Device-cloud routing accuracy | 96.1% | — |
| Very-long-context training efficiency | 4× | versus the earlier approach |
| Inference efficiency | 4.5× | from a four-layer KVCache |
This is a simplified flow based on the conference description, not on the model's technical documentation.
Beyond Spark X2.6 the company also showed the SparkAudio 1.0 speech foundation model, multimodal interaction technology for robotics, the WinWork digital-employee platform, and an AIPC device series.
What do these names mean?
This architecture rests on several terms that mean nothing without a translation.
- Edge-cloud model — an arrangement in which a small on-device model and a large cloud model work as one system. The equivalent of a local cache in front of a server.
- 4B model — a model with 4 billion parameters, small enough to fit in a phone's or laptop's memory.
- Routing — the layer deciding which model handles a given request. Like a load balancer, except the criterion is content rather than load.
- KVCache — scratch memory for already-computed context fragments, so the model does not recompute the same thing twice. iFlytek's four-layer version is said to run on fully Chinese hardware.
- Agent task completion rate — the share of multi-step tasks the model carries through to the end. Not the same as the accuracy of a single answer.
- Memory and bandwidth — the two parameters where Liu locates the main gap to NVIDIA. They determine how long a context can be loaded at all.
The first three entries describe the architecture, the next three its limits. Only together do they show why the company splits the work across two models instead of scaling one.
Why not simply wait for better hardware?
The gap is concrete and measurable. Code and agent work today can demand a million tokens of context. Liu summed up the earlier state in two phrases: it does not fit, and it cannot be trained.
Training
The company did not wait for hardware but rebuilt the training. On long tasks, compute is allocated adaptively according to task length — iFlytek puts the resulting gain in very-long-context training efficiency at roughly 4×. On the inference side the gain is 4.5×, from coordinated memory and compute plus a four-layer KVCache on Chinese hardware.
The token bill
There is a second problem that optimisation does not solve: token economics. Even a working large cloud model costs money on every request. That is why, in Liu's words, the company deliberately does not intend to build a very large cloud model.
Why the cloud is still needed
And here is the part that is easy to miss. The cloud in this architecture is not merely a fallback. Per Liu, without the cloud model's capability the on-device model cannot be trained well, because it is the cloud model that judges whether a local result meets the bar and calibrates the small model's capability boundary. iFlytek trained both layers jointly.
From that Liu draws a competitive conclusion: a vendor building only an on-device model has nothing to guide it with, and no way to decide which tasks should stay local.
Is it live yet?
Yes, it is an announced product, but the figures come solely from the vendor. The model was presented on 22 October 2026 at the 1024 developer festival.
Two scenarios were shown on stage. In the first, the model locally cross-references the weekly reports, purchase notices and project email of a robotics startup CEO and spots that the same customer was followed up twice. In the second it reviews a procurement contract, anonymising sensitive data before anything is sent.
Target devices are phones, computers, AI glasses, car cockpits and robots. The company gave no availability dates for individual hardware classes.
Why does it matter?
This is the first case where restricted access to accelerators translates not into a weaker model but into a different product architecture. If the 41.9% figure holds outside the demo, part of the agent market shifts from the cloud onto the device — and becomes less dependent on whose compute chips are available.
What comes next?
- The next target for the 4B model is 60% completion on complex agent tasks, against 41.9% today.
- Liu forecasts Chinese accelerators reaching parity within two years and is urging other vendors to train on them, noting that today only Huawei and iFlytek are doing so as a main effort.
- The company is also working on quantum computing — Liu notes that combining it with classical pretraining within three to five years would change the scale of the compute problem.
Sources
- TMTPost — 科大讯飞刘庆峰：讯飞星火X2.6全国产算力训练，押注端侧模型「绕开」封锁
- iFlytek — 2026 Global 1024 Developer Festival
