AnswersPublished27 sources
What iFLYTEK Announced About Spark X2.6 and SparkAudio 1.0
iFLYTEK announced Spark X2.6 and SparkAudio 1.0 on October 22, 2026. X2.6’s cloud model has 293B total and 30B active parameters with a 1M token context; performance and efficiency figures are company reported, not in...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- iFLYTEK announced Spark X2.6 and SparkAudio 1.0 on October 22, 2026. X2.6’s cloud model has 293B total and 30B active parameters with a 1M token context; performance and efficiency figures are company reported, not in...
- The company says Feixing 2 training for a 2T token, 256K length workload fell from 4.5 months to 1.2 months, while inference efficiency rose 4.5×.
- The X2.6 family combines a cloud model with an open source 4B on device model and a reported device cloud routing accuracy of 96.1%.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did iFLYTEK announce at its October 22, 2026 Global 1024 Developer Festival about the Spark X2.6 device-cloud model family, including i. Article summary: iFLYTEK announced the Spark X2.6 device-cloud model family and SparkAudio 1.0 at its October 22, 2026 Global 1024 Developer Festival, presenting them as advances in AI trained entirely on domestic Chinese compute. [6] Th. Topic tags: general, general web. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake numbers, clic
At its October 22, 2026 Global 1024 Developer Festival, iFLYTEK announced Spark X2.6, a device-cloud model family, alongside the SparkAudio 1.0 speech model. The company’s central claim is that it trained X2.6 entirely on domestic Chinese computing hardware and achieved coding and agent-task advances on that infrastructure. Its benchmark and efficiency figures are company-reported; the available reporting does not establish independent verification. 1
4
21
Spark X2.6: a large cloud model and a smaller on-device model
The cloud version uses a mixture-of-experts architecture with 293 billion total parameters and about 30 billion activated per token. iFLYTEK says it supports a context window of 1 million tokens. The company also said training took roughly two to three months; reporting noted that it did not disclose the specific chip model used. 1
17
23
The release also included Spark X2.6-4B, an open-source model intended for on-device use. iFLYTEK reported that it can complete 41.9% of the agent tasks it evaluated on-device, and described improvements in long-text understanding, multi-turn tool use and coordination across tools. That task-coverage figure is the company’s metric, not a general measure of model quality. 5
6
28
What iFLYTEK reported on coding and agent benchmarks
In results attributed to iFLYTEK’s published evaluations, the cloud model ranked first on SWE Pro and SWE Multilingual, and second on SWE Verified, behind DeepSeek-V4.1-Flash 1017. The company also placed Spark X2.6 in the leading group for enterprise-operations agents and the MCP-Atlas general tool-use benchmark. 17
These rankings should be read as reported test results rather than independently confirmed comparisons. The available coverage does not provide enough detail to assess whether all models were evaluated under identical conditions or to verify the results independently. 21
Domestic-compute training and inference claims
iFLYTEK said its long-task dynamic parallel training method adapts computing resources to the length of the task. On the Feixing-2 10,000-card cluster, the company reported that training 2 trillion tokens of 256K-length data took 1.2 months, down from 4.5 months—an improvement it described as roughly fourfold long-context training efficiency. 17
19
For inference, the company described heterogeneous distributed inference and storage-compute coordination as ways to address memory and bandwidth constraints on domestic chips. It also reported a four-tier KV Cache architecture, which reuses stored computation during generation: cache-hit rate rose from 63% to 95.7%, while inference efficiency reached 4.5 times its pre-optimization level. These are company-reported figures. 17
19
24
Device-cloud routing, local processing and pricing
The announced architecture can route a task to the device, the cloud or a combination of both. iFLYTEK said joint training for this routing raised accuracy from 81.2% to 96.1%. In a demonstration, the on-device model analyzed local emails and masked sensitive contract details before cloud review. That demonstration does not, by itself, establish how completely sensitive data is masked or what information is transmitted in every deployment. 5
28
35
48
Reports listed temporary API prices of RMB 0.9 per million input tokens, RMB 0.08 per million cache-hit input tokens and RMB 3.6 per million output tokens. These are promotional prices reported around the launch, not a guarantee of current pricing. 10
50
The predecessor, Spark X2.5-4B, was reported to have passed 2.56 million downloads after its September 1 release and to rank among the world’s top four on-device models. Those figures are also company-reported. 54
SparkAudio 1.0: speech and sound understanding
iFLYTEK also introduced SparkAudio 1.0, trained on domestic computing infrastructure. The announced capabilities include speech recognition, simultaneous translation and interaction, audio understanding, emotion recognition, real-time speaker separation and environmental-sound understanding. In a demonstration, it used dialogue, music and vehicle sounds in a film clip to infer the scene. 1
7
8
The company reported a 20% improvement in overall results, 15% lower response time and 60% lower usage cost. Reports said the model was being rolled out in iFLYTEK smart notebooks, Input Method, translators and iFLYTEK Tingjian products; the available evidence does not independently validate those performance improvements. 1
6
28
What the announcement establishes—and what it doesn’t
The release describes a clear product strategy: use a small model for local work and a larger cloud model for tasks that need more capacity, with routing between them. iFLYTEK presented domestic-compute training, coding and agent performance, and lower-cost long-context inference as the technical case for Spark X2.6. 1
5
17
The key caveat is evidence quality. The rankings, routing accuracy, task coverage and efficiency gains discussed here come from company-reported results in launch coverage. They are useful indicators of what iFLYTEK says the models can do, but should not be treated as independently verified benchmark findings. 17
21
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What iFLYTEK Announced About Spark X2.6 and SparkAudio 1.0"?
iFLYTEK announced Spark X2.6 and SparkAudio 1.0 on October 22, 2026. X2.6’s cloud model has 293B total and 30B active parameters with a 1M token context; performance and efficiency figures are company reported, not in...
What are the key points to validate first?
iFLYTEK announced Spark X2.6 and SparkAudio 1.0 on October 22, 2026. X2.6’s cloud model has 293B total and 30B active parameters with a 1M token context; performance and efficiency figures are company reported, not in... The company says Feixing 2 training for a 2T token, 256K length workload fell from 4.5 months to 1.2 months, while inference efficiency rose 4.5×.
What should I do next in practice?
The X2.6 family combines a cloud model with an open source 4B on device model and a reported device cloud routing accuracy of 96.1%.