iFlytek releases the Spark X2.6 cloud-device integrated large model, trained on fully domestic computing power

Zhidongxi reported on October 22 that today iFlytek released several new large model and AI application achievements, including the device-cloud integrated large model iFlytek Spark X2.6, the speech foundation model SparkAudio 1.0, the Spark embodied multimodal interaction technology, the WinWork digital employee platform, and the super AI teaching assistant iFlytek Starlight.

Its device-cloud integrated iFlytek Spark X2.6 foundation model completed a capability upgrade based on fully domestic computing power, including the cloud-side 293-A30B model and the device-side X2.6-4B model, with the latter already open-sourced.

The cloud model 293-A30B supports a million-token context, with an input price of 0.9 yuan per million tokens, an input cache hit price of 0.08 yuan per million tokens, and an output price of 3.6 yuan per million tokens. The input cache miss and output prices are lower than those of DeepSeek-V4.1-Flash-1017.

According to test results published by iFlytek, the Spark X2.6-4B model surpasses its previous generation Spark-X2.5-4B as well as the Qwen3.5-4B, Gemma4-E4B, and Qwen3.5-9B models in information search Q&A, office document processing, schedule and task management, local spreadsheet operations, cross-tool collaboration, multi-turn tool calling, and complex knowledge search.

In addition, iFlytek also officially announced several AI hardware products, such as the iFlytek AI Recording Ring, iFlytek AI Glasses, and the iFlytek AI PC series. Among them, the Lingji N9000 AIPC laptop has up to 160 TOPS of local AI computing power; the Spark Office All-in-One machine provides 2240 TOPS of AI computing power and supports private deployment of large models.

At present, iFlytek's product matrix has the iFlytek Spark foundation large model as its base, achieving applications in fields such as education, healthcare, education, law, automobiles, and robotics.

He believes that a full-stack independently controllable large model is the inevitable choice for AI security and extreme cost-effectiveness, which requires a full-stack independently controllable device-cloud integrated large model to solve. In the future, 60% of tasks will run locally, 30% of tasks will be processed through device-cloud collaboration, and 10% of tasks will be processed independently in the cloud, which is also the key reason iFlytek launched the device-cloud integrated large model X2.6 this time.

1. Trained on fully domestic computing power, first device-cloud integrated large model launched

In the benchmark test results disclosed by iFlytek, in terms of coding capability, Spark X2.6 took first place in the SWE Pro and SWE Multilingual test sets, and was second only to DeepSeek-V4.1-Flash 1017 in the SWE Verified test set; on the agent track, it is in the first tier in both enterprise operations and maintenance Agents and general tool calling MCP-Atlas; in terms of reasoning capability, Spark X2.6 is on par with MiniMax M3 in the AA-LCR v1.1 test set.

On the training side, the model broke through the long-horizon task training bottleneck of domestic chips. Taking the training of 2000B 256K-length data on the domestic ultra-large-scale intelligent computing platform Feixing-2 with ten thousand cards as an example, the training time was shortened from 4.5 months to 1.2 months, and the ultra-long context training efficiency increased to 4 times the original. Liu Qingfeng said that Spark X2.6 was trained in the past 2-3 months.

At the inference level, based on distributed inference technology for domestic heterogeneous computing power and the first four-layer KV Cache architecture on fully domestic hardware, iFlytek overcame problems such as insufficient capacity, high inference latency, and low cache hit rate in long-context scenarios. The inference efficiency of Feixing-2 domestic chips reached 4.5 times that before optimization, and the KV Cache hit rate increased from 63% to 95.7%.

In terms of the device-side model, iFlytek created the Spark X2.6-4B model. Liu Qingfeng said this is because the 4B size is very suitable for deployment on robots, automotive platforms, and AI PCs. Its previous-generation Spark X2.5-4B device-side model has been downloaded 2.56 million times since it was open-sourced on September 1, ranking among the top four device-side models globally.

To improve device-side model capabilities, iFlytek adopted device-cloud integrated training technology. Spark uses an adaptive iterative post-training algorithm that perceives device-side model capabilities, dynamically selecting teacher models and adjusting training proportions according to device-side capabilities.

Based on a "cloud + device" collaborative architecture with precise routing by the device-side model, the device-side model first determines which method should complete the task: purely on-device, purely in the cloud, or through device-cloud collaboration. A key issue here is whether device-cloud routing is accurate. iFlytek introduced a device-cloud routing collaborative training mechanism based on a critique model, raising device-cloud routing accuracy from 81.2% to 96.1%.

On top of the device-cloud integrated Spark X2.6 model, iFlytek launched two products: Spark Loomy for daily office work, helping users write reports, make PPTs, and analyze data, making daily work easier; and Spark AStudio for professional office work, helping users write code and conduct research, allowing complex work to advance autonomously.

A presenter on site played the role of the general manager of a startup company, asking Spark AStudio to help the user, without an internet connection, review recent weekly reports and emails from major regions, then analyze the business situation and organize them. After retrieving content, the model generates visualized tables, three to-do items, risks, and also provides corresponding email links.

In tasks that involve collaboration with cloud models, the model first sorts through local data, and when a task requiring cloud collaboration arises, it then initiates a collaboration request. Before that, the local model desensitizes the data. After the cloud finishes processing, the local model processes the previously desensitized information again to generate complete content.

In terms of enabling AI to move from the digital world into the physical world, iFlytek released the Spark speech foundation model SparkAudio 1.0 as well as new AI hardware including the iFlytek AI Recording Ring.

The Spark speech foundation model SparkAudio 1.0 is trained on fully domestic computing power. On site, SparkAudio 1.0 analyzed numerous sounds in a complex audio clip, including background music, male voices, mechanical operating sounds, and bicycle bells.

Behind this model are three technological innovations: a semi-supervised audio encoder that brings a 20% performance improvement, dynamic segmented end-to-end modeling that reduces response time by 15%, and dual-paradigm collaborative decoding that reduces cost by 60%.

The model has already been launched in products such as iFlytek Input 16.0, iFlytek Hearing, the iFlytek Smart Office Book, and iFlytek Translators. Among them, after the upgrade of the iFlytek Smart Office Book, it supports handwritten annotation of speaker names, with subsequent speech automatically matched; it can also distinguish speakers when people interrupt or multiple people speak over one another; and it recognizes mixed Chinese and English as well as technical terminology.

iFlytek also released a new piece of AI hardware, the iFlytek AI Recording Ring, weighing 35g and featuring a magnetic attachment design for phones. Users only need to tap the recording button on the ring to record anytime and anywhere. After recording ends, it transmits at high speed via WiFi 6, sending about 10 seconds of a 2-hour recording to the phone and converting 2 minutes of it into text, and it can also intelligently distinguish speakers, generate AI meeting minutes, extract key points and to-do items, and support multilingual simultaneous interpretation.

The device is currently on sale on e-commerce platforms, priced at 1,299 yuan.

In multimodal interaction, iFlytek has mainly improved the ability of smart terminals such as cockpits and glasses to understand needs.

The iFlytek AI Glasses have display capability, weigh 40g, support 90dB exhibition noise suppression, and pick up sound at a distance of 8 meters. This device can translate more than 120 languages, can connect to PCs and cloud agents, and supports remote command of office work on computers. The device starts at 4,299 yuan.

The iFlytek Spark Smart Cockpit SparkAuto Space 3.0 has greatly improved in multi-person interaction response rate, multi-topic interaction accuracy rate, outdoor long-range interaction distance, and success rate of complex agent tasks.

Embodied multimodal interaction for robots faces challenges such as noise-reduction recognition for robot movement in open spaces, audio source and visual spatiotemporal alignment for multi-person interaction, and multimodal interaction and action linkage coordination.

Based on this, iFlytek released the Robot Super Brain Platform V3.0, which can achieve dynamic noise reduction in complex spaces, interaction intent recognition for multiple target people, unified planning and execution of speech and actions, and vision-triggered proactive interaction.

In a "7×24-hour" AI unmanned supermarket scenario, robots can combine speech, vision, and location information to identify the needs of different customers, and even when multiple people interrupt or user needs change repeatedly, they can complete tasks such as picking up goods, delivery, and scan-code checkout.

The Robot Super Brain Platform V3.0 will be opened to developers, including the Spark Super Brain Board, multimodal interaction modules, embodied intelligence development kits, and the embodied multimodal interaction Harness.

Based on this, iFlytek launched the iFlytek Starlight WinWork platform. This platform can, by building an enterprise business ontology, accumulate an enterprise's business procedures, business data, and expert knowledge base, enabling digital employees to better understand the business; WinWork can also be embedded in enterprise business processes to support digital employees in participating in a complete work loop, from autonomously identifying tasks, decomposing and planning, and intelligent distribution to linking various business systems for execution.

In addition, iFlytek also launched the iFlytek AI PC series products based on the Spark device-cloud large model and domestic operating systems. Among them, the Lingji N9000 AIPC laptop has up to 160 TOPS of local AI computing power; the Spark Office All-in-One machine provides 2,240 TOPS of AI computing power and supports private deployment of large models.

iFlytek Starlight integrates lesson preparation, teaching research, home-school communication, and administrative assistance into the same workbench, supporting the creation of lesson plans, animations, and scenario videos, multi-role simulated roundtable discussions, standardized form filling, statistical analysis, and visual presentation. Internally, it connects to internal channels such as Zhixue.com and Xiaoxin, and externally it connects to remote control channels on multiple platforms such as WeCom, QQ, Feishu, and DingTalk.

▲Helping teachers fill out home-school communication forms

In the lesson preparation scenario, in order to help students understand the relationship between current direction and magnetic pole direction, the teacher submits a request to iFlytek Starlight, and the system generates a three-dimensional interactive resource that can be rotated for observation, and can also be combined with camera gesture operations to let students simulate the right-hand screw rule.

▲A three-dimensional interactive resource that can be rotated for observation

In the medical field, iFlytek Imaging Cloud is a global medical imaging AI service platform covering areas with more than 260 million people nationwide. The medical imaging cloud platform supports image sharing, intelligent assisted diagnosis, quality control, and report generation; iFlytek Xiaoyi is an AI health assistant by residents' side, having completed more than 220 million AI consultations cumulatively, with downloads exceeding 44 million.

The device-cloud integrated large model X2.6 launched this time is centered on finely splitting and scheduling AI task requirements, thereby balancing the contradictions among performance, cost, and data security, and providing developers and enterprises with new implementation options. Especially for industries such as education, medical care, and law that are highly sensitive to data leaving the country, this architecture can meet their needs for privatization and local deployment.

In addition, competition in today's large model industry is continuing to extend toward computing power adaptation, scenario implementation, data security, and integrated software-hardware delivery. In the future, pure foundation model players that only have model capabilities but lack industry closed loops and hardware carriers may be eliminated.