At its 2026 Global 1024 Developer Festival in Beijing on October 22, iFLYTEK launched iFLYTEK Spark X2.6, a device-cloud model family trained entirely on domestic Chinese computing power, and open-sourced a 4B on-device model for robots, cars, and AI PCs. The company also debuted the SparkAudio 1.0 speech foundation model and a new enterprise agent stack built around Agent OS and Token as a Service. All benchmark and efficiency figures below are company-reported and await independent verification.
What launched at the 1024 festival
iFLYTEK chairman Liu Qingfeng presented Spark X2.6 at the festival as the first model in China to reach a step change in coding and agent capability while trained entirely on domestic compute. The company said the model was trained over the past two to three months and is available to try on the Spark website alongside the open 4B model.
The release has three parts: a 293B-A30B cloud mixture-of-experts model with a 1 million token context window, the open-source Spark X2.6-4B on-device model with device-cloud routing, and SparkAudio 1.0 for speech recognition, translation, and understanding. A second announcement from the open platform added Agent OS, APA recorded workflow skills, the XFYUN CLI developer workbench, and Xingchen TaaS for business applications. Total model size is unchanged from X2.5, so this is an architecture and efficiency release rather than a parameter increase.
Cloud model: 293B-A30B MoE with 1M context
Spark X2.6 uses a mixture-of-experts design with 293 billion total parameters and about 30 billion active per token. In plain terms, the model keeps a large library of specialist subnetworks and calls on a small subset for each token, which keeps inference cost closer to a 30B dense model while retaining broader knowledge.
The context window extends to 1 million tokens, enough to hold a large code repository or long business documents in one pass. In practice the 293B flagship is datacenter-tier hardware, needing about 410 GB of memory at 4-bit quantization and over 820 GB at full precision, so the 4B model is the one most builders will actually run. On iFLYTEK’s own benchmarks as carried by People’s Daily Online, the flagship placed first on SWE Pro and SWE Multilingual and second on SWE Verified and ranked in the top group on enterprise operations agent tasks and the MCP-Atlas tool-use test. API pricing is published in yuan on iFLYTEK’s domestic open platform: a limited-time rate of 0.9 yuan per million input tokens, 0.08 yuan per million cached input tokens, and 3.6 yuan per million output tokens, roughly $0.13, $0.01, and $0.51 at current exchange rates. Regular rates are 1.5 yuan input and 6 yuan output, so the launch pricing is a 40% discount with no published end date, and the 4B on-device model is free for a limited time.
| Specification | iFLYTEK Spark X2.6 |
|---|---|
| Architecture | 293B total, 30B active MoE |
| Context window | 1 million tokens |
| Training compute | Domestic chips, Feixing-2 cluster |
| Coding claims | 1st SWE Pro, 1st SWE Multilingual, 2nd SWE Verified (company-reported) |
| API pricing, limited time | 0.9 input / 0.08 cached / 3.6 output yuan per million tokens (limited time, about $0.13 / $0.01 / $0.51; regular 1.5 / 6) |
Trained and served on domestic compute
Much of the presentation focused on making domestic accelerators viable for long-sequence training and inference. iFLYTEK described dynamic parallel training that works around the memory bandwidth limits of domestic chips. On its Feixing-2 cluster, training on 2,000 billion tokens of 256K-length data took 1.2 months instead of 4.5 months.
For inference, the company described distributed inference across heterogeneous domestic hardware and a four-tier KV Cache architecture built entirely on domestic hardware. The practical story is cost and control: cache hits avoid recomputation, with KV Cache hit rates reported rising from 63% to 95.7% and inference efficiency up 4.5 times. Like China’s broader push toward homegrown silicon for AI workloads, the goal is independence from foreign accelerators for both training and large-scale serving, including a domestic 10,000-card inference cluster.
Open-source Spark X2.6-4B for on-device agents
The company also open-sourced Spark X2.6-4B under Apache 2.0, aimed at robots, cars, and AI PCs, with the weights on Hugging Face, ModelScope, and Ollama. The official model card reports an average score of 63.3 across 30 evaluations covering agents, code, math, long context, and general knowledge, against 52.1 for X2.5-4B, 49.1 for Qwen3.5-9B, and 43.2 for Qwen3.5-4B. Standouts include 71.8 on BFCL-V4 function calling, 86.9 on tau2-bench, 62.8 on SWE Verified, and 64.7 on SWE Multilingual. The previous 4B model, open-sourced on September 1, is now among the four most downloaded on-device models worldwide with 2.56 million downloads, according to the company.
A router decides whether each task runs on the device, in the cloud, or across both, and local data is masked before anything is sent to the cloud. iFLYTEK says routing accuracy rose from 81.2% to 96.1% using a critique-model training setup. The pattern matches the wider shift toward small open models that handle agentic work locally, where keeping simple tasks on-device saves latency, cost, and data exposure. That same local-first logic is starting to shape consumer robots that need reliable onboard behavior.
SparkAudio 1.0 speech foundation model
iFLYTEK also debuted SparkAudio 1.0, a speech foundation model trained on domestic compute that combines recognition, translation, and semantic understanding with real-time speaker separation, sound-scene understanding, and emotion recognition. It is already running in iFLYTEK Input Method 16.0, iFLYTEK Hearing, its smart notebooks, and its translators.
The company reports 20% better accuracy, 15% lower latency, and 60% lower call costs. For accessibility and workplace tools, a single speech foundation that handles transcription plus translation and speaker separation matters more than any single benchmark number.
Agent OS, APA, CLI, and TaaS for business use
Alongside the model, iFLYTEK’s open platform launched tools for building and running agents. Agent OS handles organization and task scheduling, grouping employees, agents, and business systems into scene packages with organization-level context engineering. The company says cross-system collaboration efficiency improved by 90% and more than 75% of processes can run on their own.
APA handles execution: staff can record a screen session or describe a task in plain language, and APA packages routine work such as filling forms, checking inventory, or starting approvals into reusable skills connected to ERP and CRM systems. iFLYTEK says its procurement review system reached 97% accuracy at China Energy Investment Corporation, according to QbitAI’s festival report.
For developers, the XFYUN CLI lets a single command call more than 700 atomic AI capabilities and more than 200 agent adaptation capabilities. Xingchen TaaS sits on the Xingchen MaaS platform with model scheduling, an agent engine, knowledge engineering, tool orchestration, and runtime governance:
- Model scheduling plus an agent engine for task execution
- Knowledge engineering and tool orchestration across business systems
- Runtime governance with application packages for finance, healthcare, education, and energy
The pitch is working applications, not just tokens. That shift will feel familiar to buyers comparing AI agent platforms on real workflows.
What remains unverified and why it still matters
Every performance, download, and efficiency figure here comes from iFLYTEK. The domestic-compute training claim, benchmark placements, routing gains, and 2.56 million downloads have not been independently tested in the English-language sources available so far. Treat pricing as promotional until an end date and standard rates are published.
The constructive signal is still clear. A fully domestic training and serving stack plus an open 4B model gives builders a low-cost way to test device-cloud agents without committing to cloud APIs first. If the routing holds up, simple tasks stay local and private while hard tasks escalate, which is exactly the pattern robots, cars, and AI PCs need. Original reporting on this launch remains thin in English, and most coverage repeats the same company materials, so independent tests of long-context recall, tool-use reliability, and cache-hit rates are the next useful step.
Frequently Asked Questions
What is iFLYTEK Spark X2.6?
A 293B-A30B mixture-of-experts model family launched October 22, 2026, with a 1 million token context window, trained on domestic Chinese compute according to iFLYTEK.
What is Spark X2.6-4B?
An open-source 4B on-device model for robots, cars, and AI PCs, aimed at search, documents, scheduling, spreadsheets, and multi-step tool calls with device-cloud routing.
How much does the Spark X2.6 API cost?
API pricing is published in yuan on iFLYTEK’s domestic platform: a limited-time 0.9 yuan per million input tokens, 0.08 yuan per million cached tokens, and 3.6 yuan per million output tokens (about $0.13, $0.01, and $0.51). No dollar billing tier has been published.
What is SparkAudio 1.0?
A speech foundation model for recognition, translation, speaker separation, sound-scene understanding, and emotion recognition, already used in iFLYTEK input, hearing, notebook, and translator products.
What are Agent OS and Xingchen TaaS?
Agent OS coordinates agents across enterprise systems, while APA turns recorded or described work into reusable skills. Xingchen TaaS bundles models, agents, and industry solutions into working business applications.
