Overview
iFlytek Spark is iFlytek's model line, and speech is its deepest foundation, with the product covering dialogue, document analysis, presentation generation and coding assistance.
On 22 October 2026 at its Global 1024 Developer Day, iFlytek released Spark X2.6, a cloud-and-device model built on a 293B-A30B MoE architecture with the context window extended to 1M. Input is priced at a limited-time 0.9 yuan and output at 3.6 yuan per million tokens, the X2.6-4B on-device model was open-sourced, and three reasoning tiers were added with reasoning priced free for a limited period. The same day saw the debut of SparkAudio 1.0, a speech foundation model that folds machine translation, speech recognition and semantic understanding into one base and switches freely between Chinese and English.
Before that, Spark X2.5 shipped on 7 September 2026 with the same 293B-A30B shape and completed training and inference end to end on domestic compute; the X2.5-4B and X2.5-1.7B released on 1 September were among the first on-device models to natively support a 1 million token context.
Key Features
- Spark X2.6 flagship model: A 293B-A30B MoE with a 1M context window, priced at a limited-time 0.9 yuan per million input tokens and 3.6 yuan per million output tokens, with three reasoning tiers
- SparkAudio 1.0 foundation: Debuted on 22 October 2026 with machine translation, speech recognition and semantic understanding in one base, switching freely between Chinese and English and covering real-time speaker separation, environment understanding and emotion recognition
- Voice interaction: iFlytek speech technology feeds directly into input and dialogue, holding up on long sentences, accents and noisy rooms
- On-device deployment: X2.6-4B is open-sourced, and the earlier X2.5-4B and X2.5-1.7B natively support a 1 million token context, running on vLLM, SGLang and llama.cpp and deployable through Ollama or LM Studio
- Document analysis and slides: Upload PDF or Word files for direct Q&A, or generate a presentation from a topic
- Domestic compute: Since X2.5, training and inference run end to end on domestic compute, with continued work on long-sequence training
Use Cases
- Voice-driven dialogue and in-car scenarios
- Meeting transcription and minutes
- Office document generation and slide decks
- Chinese writing and polishing
- Coding Q&A and agent tasks
- Local deployment and incremental training on edge devices
Pros
- Speech is the long suit, with input, transcription, synthesis and a speech foundation forming one chain
- Chinese understanding and generation are steady, which suits official documents and teaching
- X2.6 extends the context to 1M, leaving more room for long documents
- The on-device models are open, running on domestic hardware and mainstream inference frameworks
- Input at 0.9 yuan and output at 3.6 yuan per million tokens keeps cost predictable
- Core features are free to use
Pricing
Core features are free. Spark X2.6 bills by token at a limited-time 0.9 yuan per million input tokens and 3.6 yuan per million output tokens, with reasoning priced free for a limited period across the three tiers. The X2.6-4B on-device model is open-sourced for self-hosting. The professional plan adds larger quotas and advanced features, with current terms listed on the official site.
Summary
iFlytek Spark has held a clear position: speech is the entrance and the model is the foundation. X2.6 brings the parameters and context to 293B-A30B and 1M, prices input at 0.9 yuan and output at 3.6 yuan per million tokens, and leaves the open X2.6-4B to cover local deployment, while SparkAudio 1.0 closes the loop on the speech side. For Chinese voice interaction, meeting transcription and deployment on domestic compute, it is one of the few fully self-developed options in the country.
Version History
- Spark X2.6 and SparkAudio 1.0 (2026-10-22): iFlytek released Spark X2.6 at its 1024 Developer Day, a 293B-A30B MoE with a 1M context window priced at a limited-time 0.9 yuan per million input tokens and 3.6 yuan per million output tokens, alongside an open-sourced X2.6-4B on-device model, three reasoning tiers and the first SparkAudio 1.0 speech foundation model.
- Spark X2.5 released with a focus on coding and agents (2026-09-07): iFlytek released Spark X2.5, a 293B-A30B MoE model focused on coding and agent capability alongside stronger math and comprehension. Training and inference run end to end on domestic compute, continuing work on long-sequence training and inference on domestic hardware. It is live on the iFlytek Open Platform.
- Spark X2.5-4B and 1.7B on-device models open-sourced (2026-09-01): iFlytek subsidiary CiYuan XingHuo released and open-sourced Spark X2.5-4B and X2.5-1.7B on-device models, the first on-device models to natively support a context window of up to 1 million tokens. Built on a hybrid attention architecture and tuned for agents, code, math and instruction following, they run on NVIDIA, Huawei, Hygon and Houmo hardware, work with vLLM, SGLang and llama.cpp, deploy through Ollama and LM Studio, and support incremental training with LLaMA-Factory.