Spark X2.6 Operational

Spark X2.6 (Spark-X2.6) is an edge-cloud integrated foundation model launched by iFlytek. Trained on fully domestic computing power, it is a key move for China's AI industry in security and cost-effectiveness. The cloud-based 293-A30B supports a million-token context, with input priced at RMB 0.9 per million tokens, lower than DeepSeek-V4.1-Flash. The edge-side Spark X2.6-4B is open source and excels at agent tasks such as office documents and tool collaboration. Based on task difficulty and data security requirements, the system automatically assigns tasks across edge-only, cloud-only, and edge-cloud collaboration modes, keeping sensitive data on the device while handling simple tasks at low cost. Its coding and agent capabilities rank among the top tier, and it has already been deployed in office products such as Loomy and AStudio.

Million-token context understanding: Cloud-based 293-A30B supports a million-token context and can process extremely long documents and complex multi-turn tasks.
Code generation and repair: It ranks first in code benchmarks such as SWE Pro and SWE Multilingual, reaching an international leading level.
Agent task execution: It supports multi-turn tool calls and cross-tool collaboration. Enterprise operations agents and MCP-Atlas tool calling rank among the top tier.
Local edge intelligence: X2.6-4B can complete office tasks such as email analysis, weekly report summaries, and business analysis locally and offline, keeping data on the device.
Edge-cloud collaborative office work: Spark Loomy and AStudio are built in, covering scenarios such as report writing, PPT creation, coding, and data analysis.

Follow WeChat and reply “开源” to join the AI open-source project community.
Long-range training on fully domestic computing power: Based on the Feixing-2 10,000-GPU domestic intelligent computing platform, it uses dynamic parallel training for long-range tasks and adaptively allocates computing resources according to task length. It reduces the training time for 2,000B data at a 256K length from 4.5 months to 1.2 months, improving long-context training efficiency by approximately 4 times and achieving complete independent control from training through inference on domestic chips.
Four-layer distributed inference with KV Cache: To address the limited memory and bandwidth of domestic chips, the system builds the first four-layer KV Cache architecture based on fully domestic hardware, together with adaptive compute-storage collaborative inference technology. This greatly improves cache reuse and computing efficiency, increasing inference efficiency by 4.5 times and cache hit rate from 63% to 95.7%, resolving the bottlenecks that make long contexts difficult to store and slow to compute.
Edge-cloud integrated collaboration architecture: It uses a routing mechanism in which the edge evaluates tasks and the cloud collaborates. The edge model first assesses task difficulty and data sensitivity, then selects edge-only, cloud-only, or edge-cloud collaboration. Combined with routing and collaborative training based on an evaluation model, routing accuracy rises from 81.2% to 96.1%. Data is desensitized before being uploaded to the cloud, balancing security and cost.
Distillation of edge-side model capabilities: X2.6-4B acquires its capabilities through edge-cloud integrated training. It dynamically selects teacher models and adjusts training weights based on the current capabilities of the edge model, while using an adaptive iterative post-training algorithm. As a result, its edge agent task coverage reaches 41.9%, surpassing comparable Qwen and Gemma models on multiple office tasks.

Define the use case: Based on task difficulty and data sensitivity, decide whether to use the cloud API, local edge deployment, or both collaboratively.
Cloud access: Visit the iFlytek Open Platform at https://maas.xfyun.cn/modelSquare/base/2628905721051136?sort=0&userSelectCategory=%255B%255D and call the 293-A30B API. Input is priced at RMB 0.9 per million tokens, making it suitable for complex reasoning and long-document tasks.
Open-source edge deployment: Download X2.6-4B from HuggingFace (https://huggingface.co/XHToken/Spark-X2.6-4B) and deploy it on an AI PC, vehicle, or robot platform.
Direct use in office scenarios: For everyday use, choose Spark Loomy or Spark AStudio, which support offline processing of local files.
Configure edge-cloud collaboration: Enable edge-cloud routing so the system automatically determines whether tasks should be completed locally, in the cloud, or collaboratively. Sensitive data can be configured for edge-only processing.
Enterprise private deployment: Enterprises with strict data security requirements can use the Spark office all-in-one machine for localized private deployment.

Independent control through fully domestic computing power: It is the first to achieve a complete training-to-inference loop based on domestic chips, making it a truly full-stack independently controlled foundation model.
Million-token long context: Cloud-based 293-A30B supports million-token input and can process extremely long documents and complex tasks in one pass.
Intelligent edge-cloud scheduling: It automatically selects edge-only, cloud-only, or edge-cloud collaboration according to task difficulty and data security requirements, with routing accuracy reaching 96.1%.
Data stays on the device: Sensitive tasks can be completed locally and offline. Combined with desensitization mechanisms, this addresses urgent data compliance needs in education, healthcare, legal services, and other industries.
Leading coding and agent capabilities: It ranks first in SWE Pro and SWE Multilingual coding tests, while its agent capabilities rank among the top tier.
Strong, open-source edge capabilities: X2.6-4B is open source and deployable. Its edge agent task coverage reaches 41.9%, surpassing comparable Qwen and Gemma models.

GitHub repository: https://github.com/XHToken/Spark-X2.6
HuggingFace model collection: https://huggingface.co/collections/abenzerps/spark-x26

Enterprise-grade edge-cloud collaborative AI services: Through a cloud-and-edge routing architecture, simple requests are handled instantly on the edge while complex tasks are forwarded to the cloud foundation model, substantially reducing cloud computing costs while maintaining the user experience.
Intelligent office work: Loomy and AStudio handle tasks such as report writing, PPT creation, data analysis, and coding, while supporting offline processing of local emails and documents.
Enterprise business analysis: The edge model reads local weekly reports and emails, automatically summarizes business conditions, generates visual tables, and flags risks and pending tasks.
Code development: Leveraging its first-place performance in SWE Pro and SWE Multilingual, it assists developers with code generation, repair, and maintenance of multilingual projects.