iFlytek subsidiary open-sources Spark X2.5 on-device models with million-token context
The AMW Read
iFlytek's subsidiary extends its foundation-model lineup with an explicit open-weight, long-context on-device release aimed at agentic edge task execution, updating a known CN player without resolving an open debate.
iFlytek subsidiary open-sources Spark X2.5 on-device models with million-token context
iFlytek's wholly owned subsidiary Ciyuan Xinghuo (词元星火) open-sourced two on-device language models on September 1: Spark X2.5-4B and Spark X2.5-1.7B, which the company describes as the industry's first on-device open-source models to natively support a 1-million-token context window. On a Domux smart-home benchmark, the smaller 1.7B model reportedly reached 90.3% end-to-end accuracy on device-control instructions with an average 0.85-second response time. Both models were trained entirely on domestic Chinese compute platforms and support Nvidia, Huawei, Hygon, and Houmo hardware, plus vLLM, SGLang, and llama.cpp inference frameworks; weights and code are published on GitHub and Hugging Face, with API access offered free for a limited time on iFlytek's Spark MaaS platform.
The release targets a structural gap in consumer AI hardware: devices that can converse fluently but cannot execute multi-step tasks because they lack persistent memory of conditions and rules, reliable offline reasoning, and tool-calling ability. iFlytek is positioning long-context plus agentic tool use as the next competitive axis for smart-home, robotics, and wearable products, as the category shifts from voice-recognition accuracy toward task-completion reliability. Per the AI Market Watch index, iFlytek logged zero pipeline-tracked items in the past 90 days against one in the prior 90 (name-matched, pipeline-ingested sources only), making this open-source release the company's first tracked move in months.
For builders, the practical draw is deployment breadth: fine-tuning support via LLaMA-Factory and quick setup through Ollama or LM Studio lower the integration cost for smart-home and robotics OEMs building offline-capable assistants, particularly where cloud dependency raises latency or privacy concerns. For investors, the release signals that CN on-device model competition is shifting toward context length and agentic execution as differentiators rather than raw chat quality, worth tracking against rival open-weight releases from other domestic labs.
