iFlytek Spark X1.5

A deep reasoning model on a domestic stack

iFlytek released Spark X1.5 on 6 November 2025 at the iFlytek 2025 Global 1024 Developer Festival^[1]. It is a deep reasoning model, meaning it is built to work through a problem in several steps before giving an answer rather than producing a response in a single pass^[1]. It runs on a domestic computing platform built with Huawei^[1]. The model and the hardware underneath it are therefore presented as a matched pair: the software was developed for, and tuned to, computing hardware that is not the mainstream international stack. That matters because a reasoning model spends much of its compute on intermediate steps, so the cost of serving it is set as much by the hardware and software stack as by the model itself.

Sparse mixture of experts

Spark X1.5 uses a mixture-of-experts (MoE) architecture with 293 billion total parameters^[1]. In an MoE model the network is split into many separate sub-networks called experts, and a router decides, token by token, which few of them to use. The total parameter count therefore describes everything the model stores, not everything it runs at once. During inference, 3 billion parameters are active^[1]. That gap is the point of the design: computation per token follows the active count, while capacity follows the total, so a 293-billion-parameter model can be served at a cost closer to that of a much smaller dense model. The trade-off is that routing has to be learned and kept balanced, which makes training harder than training a dense network of the same active size.

Efficiency

iFlytek said inference efficiency doubled compared with the previous generation, a figure Liu Qingfeng presented^[1]. The company also said it had greatly improved full-chain MoE training efficiency on domestic compute^[1] — the harder half of the problem, since training a sparse model means keeping the routing and the experts in balance across an entire run, not just at inference time.

Abilities and benchmarks

iFlytek benchmarked six core abilities against leading international models, including language understanding, logical reasoning, mathematics and coding^[1]. The model supports more than 130 languages and covers 60 countries and regions^[1]. iFlytek put its overall capability at 95% of GPT-5 and said it was above 93% of international benchmark performance^[2]. Those figures are the company's own comparisons against named international systems; they describe where iFlytek placed the model at launch rather than an independent ranking.

Memory, voice and agents

Spark X1.5 added personalised memory and voice cloning^[1]. Alongside the model, iFlytek launched the Astron agent platform with native RPA (robotic process automation) support, offering 8 industry agents^[1]. RPA is the layer of software that drives existing applications the way a person would — filling in forms, moving records between systems — so an agent platform with native RPA is aimed at tasks that end in an actual change inside a business system, not just a paragraph of generated text. Personalised memory works in the same direction: it lets a model carry user-specific context across sessions instead of starting from nothing each time.

Place in the line

Spark X1.5 was succeeded by Spark X2 in February 2026^[3], roughly three months after launch — a short interval that shows how quickly the model line was being revised.