iFlytek Spark X2.5

What it is

iFlytek Spark X2.5 is a large language model released on 7 September 2026^[1][2]. It uses a mixture-of-experts (MoE) architecture with 293 billion total parameters and roughly 30 billion active parameters, a size conventionally written 293B-A30B^[1][2]. iFlytek describes it as a multilingual text-generation model built on a fully domestic platform^[1].

Why the parameter count comes in two numbers

In a dense model, every parameter is used for every token the model processes. In a mixture-of-experts model the network is divided into many separate "expert" sub-networks, and a routing step sends each token to only a few of them. The total parameter count therefore measures capacity — how much the model can store — while the active count measures the computation a single token actually costs. A 293B-A30B model carries the parameter budget of a much larger dense model but, in principle, runs at something closer to the cost of a 30-billion-parameter one. That is why the ratio is quoted alongside the release rather than the total alone.

Trained and served on domestic computing power

iFlytek states that both training and inference for X2.5 were carried out fully on domestic computing power^[2]. Training is the long offline phase in which a model's weights are adjusted over large amounts of text; inference is the everyday serving of user requests. Running both on domestic hardware means the whole lifecycle of the model, not just a demonstration run, sat on that stack.

Capabilities

The model supports a context of 256K tokens and more than 200 languages^[1]. The context window is the amount of text — the prompt plus whatever material the model is reading — that it can hold at once, so 256K tokens is enough for long documents or sizable code bases in a single request. Coverage of the release said X2.5 keeps strong results on general tasks such as language understanding, question answering and reasoning^[1]. The stated focus of the release was stronger coding and agent capabilities, alongside improvements in mathematics and understanding^[2]. An "agent" here means a model that carries out multi-step work — calling tools and acting across several turns — rather than answering one question and stopping.

Smaller siblings on the device side

Ahead of the main release, iFlytek open-sourced two on-device models, X2.5-4B and X2.5-1.7B, supporting contexts of up to 1 million tokens^[2]. On-device models run locally on a phone or a personal computer instead of in a data centre; the small parameter counts are what make that practical, and the long context is what lets such a model work through a lengthy local document.

Availability and price

X2.5 is available on the iFlytek open platform^[2]. On the Xingchen MaaS platform it is priced at RMB 1.6 per million input tokens, RMB 0.24 per million on a cache hit, and RMB 6 per million output tokens^[1]. A cache hit means the request reuses a prefix the service has already processed, so the repeated part is billed at a far lower rate — which matters for applications that send the same long instruction or document with every call.

Where it sits

Spark X2.5 was succeeded by Spark X2.6 in October 2026^[3], which places it as one step in a rapid sequence of releases rather than an endpoint.

^[1]: 同花顺财经，《科大讯飞星火 X2.5 模型正式发布：293B-A30B MoE，全国产平台训练》 ^[2]: 新京报，《科大讯飞发布星火X2.5大模型》 ^[3]: 凤凰网科技，《科大讯飞发布端云一体讯飞星火X2.6大模型 让敏感数据留在端，把复杂任务交给云》