iFlytek Spark X2.6: 293B Model Trained on Chinese Chips

iFlytek Spark X2.6 is the Chinese AI company's new flagship model, launched on October 22 at its 1024 Developer Festival in Beijing with a 1 million token context window and a loud claim: it was trained entirely on Chinese-made chips. The flagship is a 293 billion parameter mixture-of-experts model aimed at coding and agents, while a small 4B sibling ships with open weights.

What actually happened

At the iFlytek 1024 Global Developer Festival in Beijing on Thursday, the company unveiled Spark-X2.6. The Global Times describes it as an "end-cloud integrated" model, meaning it is designed to work across devices and the cloud, with breakthroughs in coding and agentic tasks.

Chairman Liu Qingfeng put the chip story front and center. "Spark-X2.6 has taken the lead in breaking through the efficiency bottleneck of long-horizon task training on domestic chips, truly achieving independence and controllability with homegrown computing power," he said, per the Global Times. AGI Pulse reports he also claimed Spark is the only mainstream Chinese model not trained on NVIDIA GPUs.

Liu argued the model's real-world performance in sectors like state-owned enterprises, education, healthcare and courts beats its public benchmark scores. That is a bold claim to make without publishing the benchmarks.

What iFlytek Spark X2.6 actually ships

Here is where headlines get messy. The Global Times wrote that iFlytek "released and open-sourced" the model. AGI Pulse, which checked the release materials, says only the small Spark-X2.6-4B is open, under Apache 2.0, with a 1 million token context and three reasoning levels. The 293B flagship stays closed and is sold through iFlytek's API.

So if you want to tinker locally, the 4B version is the one you can download. It is listed on Hugging Face under iFlytek's XHToken account and on Ollama as SparkLLM/Spark-X2.6-4B. The flagship is an API product.

On price, AGI Pulse puts the promo rates close to MiniMax M3.1 Flash at $0.15 and $0.60, and below DeepSeek V4.1 Flash at $0.30 and $1.20 per million tokens. The difference is that those rivals publish benchmarks, and iFlytek Spark X2.6 does not, at least not for the flagship.

Cheap, long context, sovereign chips. The missing piece is independent evaluation.

Why the chip angle matters

AGI Pulse notes iFlytek has trained Spark on Huawei Ascend clusters since 2023. The Global Times also reported that a team led by the Shenzhen Loop Area Institute used Huawei's Ascend 910C chips to complete post-training for DeepSeek-V4-Pro, which it framed as a step from inference toward full training on homegrown hardware.

For government and state-linked buyers in China, a stack free of US export controls is a selling point by itself. For developers elsewhere, it matters mostly if the model is actually good, and that is still unproven. AGI Pulse flags the "only all-domestic" claim and the training scale as unverified.

Want the competitive context? Our MiniMax M3.1 Flash piece and DeepSeek V4.1 Flash breakdown cover the two cheap models AGI Pulse compares it with, and our StepFun Step 5 story covers another Chinese lab chasing agents.

What to do as a reader

If you build with LLMs, try the open 4B model first. It costs nothing to test and tells you something about the family. For the flagship, start on iFlytek's Spark API page and run your own tasks, since there are no public flagship benchmarks to lean on.

If you follow AI as a narrative, the takeaway is that Chinese labs are pushing hard on domestic chips and low prices. That pressure tends to push API prices down for everyone.

If you see tokens or "Spark" coins riding this headline, ignore them. iFlytek Spark X2.6 is a model, not a crypto project, and nothing here is investment advice. Watch for independent evals in the coming weeks, ser. That is when we will know if the hype holds.

Sources: Global Times report and AGI Pulse release analysis.