More

    iFLYTEK Open Sources Spark X2.6, a 293B-Parameter Model Trained on Domestic Chips

    iFLYTEK released and open-sourced Spark X2.6 at its 1024 Global Developer Festival in Beijing on October 22, 2026, describing it as an end-cloud unified large model trained entirely on domestic computing power.

    The cloud model uses a 293B-A30B mixture-of-experts design, meaning 293 billion total parameters with roughly 30 billion active per token. It supports a 1-million-token context window, and iFLYTEK says it ranked first on SWE Pro and SWE Multilingual and second on SWE Verified on its own benchmarks.

    What is actually open

    iFLYTEK also open-sourced the on-device Spark X2.6-4B under Apache 2.0 for robots, cars and AI PCs, and debuted SparkAudio 1.0, a speech foundation model combining recognition, translation, speaker separation and emotion recognition. Launch pricing was listed at 0.9 yuan per million input tokens, 0.08 yuan per million cached input tokens and 3.6 yuan per million output tokens.

    The distinction worth holding onto is between the closed flagship and the open sibling. The 293B cloud model is closed; the 4B on-device model is the one you can actually download and run. That matters if you are evaluating this for a project rather than reading the headline.

    What is vendor-stated

    The claim that the model was trained entirely on domestic computing power, and all the benchmark rankings, are the company’s own statements. Coverage of the launch comes from state-linked Chinese outlets and secondary aggregators; we did not locate an iFLYTEK English-language press release or an independent model card, and no second independent outlet has confirmed the domestic-compute claim.

    That is why we are reporting this as a company announcement rather than an independently verified result. The open-sourced 4B model is the verifiable part of the story; the flagship’s training claim is not.

    For contrast, our report on OpenAI models arriving on Google Cloud covers a closed model distribution deal announced the previous day.

    What we did not do

    We did not run, benchmark or download either model, and we did not verify the 1-million-token context window or any of the benchmark rankings. The pricing figures are iFLYTEK’s published launch rates and may change.

    Rows of server racks in a data centre, illustrating compute infrastructure.
    Server racks in a data centre. Photo: Carl Lender, Wikimedia Commons (CC BY 2.0).

    Sources

    Latest articles

    spot_imgspot_img

    Related articles

    Leave a reply

    Please enter your comment!
    Please enter your name here

    spot_imgspot_img
    HF

    Honest Fred AI Assistant

    Always here to help

    Images processed securely, not permanently stored. Don't share passwords or sensitive info.