iFlytek releases the Spark X2.6 series models and open-sources the 4B on-device version

iFlytek Unveils Spark X2.6 Series Models! iFlytek's 1024 Developer Conference Goes Wild Again

Sandwich 1 hour ago iFlytek AI fully upgraded.

From the earliest ChatGPT to the later collective explosion of domestic large models, vendors announce new breakthroughs in their models every few days. But what Xiaolei cares more about is: besides chatting with us, what else can these AIs actually do?

On the afternoon of October 22, Leitech was invited to attend the 2026 iFlytek Global 1024 Developer Festival. This year's conference theme is "Liberate Productivity, Unleash Imagination," and iFlytek brought out in one go the device-cloud integrated Spark X2.6, the speech foundation model SparkAudio 1.0, the all-modal human-like interaction cockpit, and the Xingfan WinWork digital employee platform.

(Image source: iFlytek)

The highlight today, naturally, is still iFlytek Spark X2.6.

You may recall that since releasing the first generation of the Spark cognitive large model in 2023, iFlytek has consistently emphasized domestic computing power, and Spark X2.6 is the first to achieve breakthroughs in coding and agent capabilities based on fully domestic computing power.

iFlytek Chairman Liu Qingfeng said that currently only Spark in the industry is trained on fully domestic computing power; all others are trained on N cards, and at most later migrated to run on domestic computing clusters, and iFlytek's size is the smallest among them.

On this basis, the Spark X2.6-4B open-source on-device large model was officially released, taking the lead in achieving a leap in on-device large model agent capabilities, and based on this model brought a brand-new device-cloud integrated collaborative architecture, raising the collaboration efficiency between the local device model and the cloud large model to 96.1%.

(Image source: iFlytek)

The iFlytek Spark X2.6 series models are now available for trial on the iFlytek Spark official website, with pricing: input limited-time 0.9 yuan per million tokens; input cache hit limited-time 0.08 yuan per million tokens; output limited-time 3.6 yuan per million tokens.

This price point belongs to the low-price tier in the current domestic large model API market, somewhat cheaper than Zhipu and Doubao, and only slightly more expensive than Xiaomi MiMo and DeepSeek's Flash models, making it quite cost-effective.

(Image source: Leitech self-made)

Besides large models, intelligent speech is also an old specialty iFlytek cannot avoid.

The officially released SparkAudio 1.0 integrates capabilities such as machine translation, speech recognition, and semantic understanding, significantly improving basic speech capabilities.

iFlytek demonstrated the model's performance in meetings, and it is not hard to see that the model performs well in real-time speaker separation, environment understanding, audio understanding, and emotion recognition. iFlytek hopes to let AI understand both content and speaking style at the same time, pushing artificial intelligence toward applications in the physical world.

(Image source: iFlytek)

Currently SparkAudio 1.0 is fully launched on iFlytek Input Method 16.0, iFlytek Hearing, iFlytek Office Notebook, and iFlytek Translator, supporting arbitrary switching between Chinese and English and real-time speaker recognition.

By the way, an iFlytek AI recording ring was also released on site. It can not only record but also transcribe; a 2-hour recording takes only 2 minutes to transcribe, and it can even serve as a phone stand. The official website price is 1,299 yuan.

(Image source: iFlytek)

Let's look at cars next.

Many new cars today can control the air conditioning, windows, and navigation by voice, and some models can even understand several commands at once. But when multiple people speak at the same time, or when the user speaks vaguely, the experience may not be so smooth.

Today, iFlytek debuted the all-modal human-like interaction cockpit — the new generation Spark Smart Cockpit SparkAuto Space 3.0, hoping to combine information such as voice and vision to let the car system more naturally understand the needs of passengers in the car, pushing smart cockpits from feature stacking toward reshaping the human-vehicle relationship.

(Image source: iFlytek)

In terms of functionality, the biggest feature of this cockpit is support for simultaneous interaction by multiple people, and it can also remember user habits, with a 95% accuracy rate in understanding complex travel intentions.

By the way, iFlytek also released the enterprise-oriented Xingfan WinWork digital employee platform this time.

(Image source: iFlytek)

WinWork hopes to let enterprises configure digital employees according to different positions, letting AI participate in material organization, information query, and business process handling. In the future, those tasks that require repeated operations every day may be handed over to digital employees, so that real employees only need to be responsible for checking and handling more complex problems.

Looking back at the 1024 Developer Festivals of recent years, iFlytek's progress is quite obvious.

From the release of the first-generation Spark large model in 2023 to today's launch of X2.6, the speech large model, and digital employees, in three years iFlytek has brought AI into multiple fields such as education, healthcare, automobiles, and office work.

This is of course a good direction. After all, ordinary consumers will not study model parameters every day, nor will they buy a new device just because a certain benchmark score leads. When we open our computers or sit in our cars and can naturally let AI solve problems, only then will these new technologies truly be integrated into life.

Images in this article are from: 123RF licensed image library Infringement report Return Share Related articles All comments