# What Is Spark X2.6? iFlytek’s New AI Model, Explained

URL: https://whatisagenticai.net/blog/what-is-spark-x2-6/
Published: 2026-10-22
Publisher: Little Tough LLC, maker of AI Agent Assistant: Agentic AI (https://apps.apple.com/us/app/ai-agent-assistant-agentic-ai/id6761501260)

Spark X2.6 is iFlytek’s new AI model, launched October 22, 2026 and trained on Chinese chips. Specs, reported price, and how to try the free open 4B version.

Spark X2.6 is the newest AI model from iFlytek, the Chinese company best known for speech recognition and voice typing. It was unveiled on October 22, 2026, at iFlytek’s yearly “1024” developer festival in Beijing, and its main selling point is not a benchmark. It is where it was built: entirely on Chinese computer chips.

If you saw “iFlytek Spark X2.6,” “Spark-X2.6” or “Xinghuo X2.6” in the news and wondered what it is, whether it is open source, and whether you can try it, here is the plain-English answer.

**TL;DR:** Spark X2.6 is iFlytek’s new large language model, launched October 22, 2026. The big version is a mixture-of-experts model with 293 billion parameters (30 billion active per word) and a 1 million token context window, trained and served on Chinese hardware. iFlytek also released a small open model, Spark-X2.6-4B, under the free Apache 2.0 license on Hugging Face, ModelScope and Ollama. iFlytek claims gains in coding and agent tasks but has not published independent benchmark numbers.

## What is Spark X2.6?

Spark (Xinghuo, 星火, in Chinese) is iFlytek’s family of AI models, a rival to ChatGPT, Qwen and DeepSeek inside China. X2.6 is the version that follows Spark X2.5, which came out in early September 2026.

iFlytek calls X2.6 an “end-cloud integrated” model. In normal words: a large model runs in the cloud, a small model runs on the device (a phone, PC or car), and the two are built to work together. The small one handles quick tasks close to you, and the big one takes over when the job gets hard.

Chairman Liu Qingfeng framed the launch around self-reliance, saying it achieves “independence and controllability with homegrown computing power,” per the Global Times.

## Spark X2.6 at a glance

| Fact | Spark X2.6 (cloud) | Spark-X2.6-4B (open) |
| --- | --- | --- |
| Made by | iFlytek (China) | iFlytek’s SparkLLM team |
| Released | October 22, 2026 | October 21 to 22, 2026 |
| Size | 293B total, 30B active (mixture of experts) | 4B |
| Context window | Up to 1M tokens | Up to about 1M tokens in the official server setup |
| Weights | Not published as of October 22 | Free download, Apache 2.0 |
| Reported price | 0.9 yuan input, 0.08 yuan cached input, 3.6 yuan output per million tokens | Free inference for a limited time |
| Main pitch | Coding and agents on Chinese chips | Agents and tool use on devices |

The prices come from IT Home’s launch report. We could not yet confirm them on iFlytek’s own pricing page, so treat them as reported.

## Why “domestic computing power” is the headline

Most top AI models are trained on Nvidia GPUs. US export rules limit which Nvidia chips Chinese companies can buy, so Chinese labs are racing to prove they can do the same work on local hardware, mostly Huawei’s Ascend chips. iFlytek says X2.6 was trained on fully domestic infrastructure and runs on a 10,000-card Chinese inference cluster.

That is hard because domestic chips have less memory bandwidth. According to IT Home’s summary, iFlytek built new parallel training and distributed inference methods to work around it, especially for long tasks with very long inputs, which is exactly what agents need. Other Chinese labs are making the same move: DeepSeek’s open-source [DeepInfer engine](https://whatisagenticai.net/blog/what-is-deepinfer/) also treats Huawei Ascend as a first-class option.

## Is Spark X2.6 open source?

Partly. The open piece is **Spark-X2.6-4B**, a small model with 4 billion parameters. Its model card says it keeps the same design as Spark-X2.5-4B and gets its gains from better post-training. It lists three “thinking levels” (none, low and high), so an app can trade a little accuracy for speed. The license is Apache 2.0, which allows commercial use.

The large 293B model is a different story. We found no public weights for it on launch day; it is available as a service from iFlytek. If that changes, we will update this page.

## How to try Spark X2.6

- **On the web:** iFlytek says the X2.6 series can be tried on its official Spark website. The site is built mainly for users in China.

- **Through the API:** developers can reach it through iFlytek’s platform, at the prices reported above.

- **On your own computer:** the 4B model runs with Ollama (ollama run SparkLLM/Spark-X2.6-4B), LM Studio, llama.cpp, vLLM, SGLang and Hugging Face Transformers. There is also an MLX build for Apple silicon Macs.

## Spark X2.6 vs Spark X2.5

|  | Spark X2.5 | Spark X2.6 |
| --- | --- | --- |
| Launch | September 2026 | October 22, 2026 |
| Big model size | 293B total, 30B active | 293B total, 30B active |
| Context window | 256K tokens | 1M tokens |
| Open small models | X2.5-4B and X2.5-1.7B | X2.6-4B |
| Focus | Coding and agents | Coding and agents, device plus cloud |

So the jump is less about size and more about a four times longer memory, better agent skills and the new device-plus-cloud design. iFlytek says the new 4B model beats the old 4B on every benchmark it tested and matches some 9B models, but the scores are only published as images and nobody has checked them independently yet.

## What this means for you

If you live outside China, you will probably not use Spark X2.6 directly. What it shows is the bigger trend: AI makers are racing to build models that act, not just chat. Coding, tool use and long tasks are what every launch now leads with, from [MiniMax M3.1 Flash](https://whatisagenticai.net/blog/what-is-minimax-m3-1-flash/) to [Mistral Large 4](https://whatisagenticai.net/blog/what-is-mistral-large-4/).

You do not need to track every new model to get that benefit. That is the idea behind [agentic AI](https://whatisagenticai.net/what-is-agentic-ai/): you describe the job, and agents plan it, do it and check it. **Agentic AI** gives you a team of up to four expert agents on iPhone and iPad, with no account needed, so you can skip the model shopping.

## Quick answers

**What is iFlytek Spark X2.6?** iFlytek’s newest large language model, launched October 22, 2026. It has 293 billion parameters (30 billion active), a 1 million token context window and was trained on Chinese chips.

**Is Spark X2.6 open source?** The small Spark-X2.6-4B model is open under Apache 2.0. The large 293B model had no public weights as of launch day.

**Where can I download Spark-X2.6-4B?** On Hugging Face and ModelScope in the SparkLLM collection, or with Ollama using ollama run SparkLLM/Spark-X2.6-4B.

**How much does Spark X2.6 cost?** IT Home reports 0.9 yuan per million input tokens and 3.6 yuan per million output tokens, with limited-time free inference for the 4B model. iFlytek’s own price page has not confirmed this yet.

**Is Spark X2.6 better than DeepSeek or Qwen?** iFlytek says it is comparable to leading models on several benchmarks, but it has not published full scores, and no independent tests are out yet.

**Who makes Spark AI?** iFlytek (科大讯飞), a listed Chinese company based in Hefei that started in speech recognition.

New to words like “mixture of experts” or “context window”? Our [agentic AI glossary](https://whatisagenticai.net/agentic-ai-glossary/) explains them in plain English.

**Sources:** [Global Times](https://www.globaltimes.cn/page/202610/1372798.shtml) · [Spark-X2.6 on GitHub](https://github.com/XHToken/Spark-X2.6) · [Spark-X2.6-4B on Hugging Face](https://huggingface.co/XHToken/Spark-X2.6-4B) · [IT Home report via KuCoin News](https://www.kucoin.com/news/flash/iflytek-launches-x2-6-end-cloud-ai-model-based-on-fully-domestic-compute-power) · [10jqka on Spark X2.5](https://stock.10jqka.com.cn/20260907/c679649499.shtml)