iFlytek releases Spark X2.5

What was released

On 7 September 2026, iFlytek released a new version of its Spark model, Spark X2.5 ^[1]. The model is a mixture-of-experts (MoE) system labelled 293B-A30B: 293 billion parameters in total, of which about 30 billion are active for each token the model processes ^[1]. Both training and serving run on domestic compute ^[1].

The MoE label matters for how a model of this size is used. In a dense model, every parameter is involved in every token; in a mixture-of-experts model, a routing step picks a small subset of specialised sub-networks for each token, and only those run. That is how a model with a large total parameter count can be served at the cost of a much smaller active count — roughly what the "293B-A30B" notation records.

Code and agents in focus

iFlytek's release emphasises code and agent capabilities ^[1]. Code here means generating and working with program text; agent capability generally means the model can carry out a multi-step task, deciding what to do next and calling external tools, rather than only producing a single reply. The two are often grouped together in recent releases because writing code and completing agent tasks both put pressure on long, structured outputs and on keeping track of earlier steps.

Small open models with long context

Alongside the main model, iFlytek open-sourced two on-device models, X2.5-4B and X2.5-1.7B, which support up to 1 million tokens of context ^[1]. The "B" suffix conventionally denotes billions of parameters, so these are far smaller than the 293B model and sized to run locally rather than in a data centre. A context window of up to 1 million tokens describes how much text the model can hold in view at once — a conversation, a set of documents or a codebase it is reasoning over — before older material has to be dropped.

Price and availability

The API is priced at RMB 1.6 per million input tokens and RMB 6 per million output tokens ^[2]. Output is charged at a higher rate than input, since generating text takes more computation than reading a prompt. The model is available on iFlytek's Xingchen MaaS platform ^[2] — a model-as-a-service offering, in which developers reach models through an API instead of deploying and maintaining them themselves.

General capability

The release states that the model keeps strong results on general tasks such as language understanding, question answering and reasoning ^[2]. These are the everyday jobs a general-purpose model is expected to handle, and holding them in place matters when a release is tilted toward code and agent use: a model that gains on specialised tasks but loses ground on ordinary understanding is harder to use as a general assistant.

Why it matters

Two things stand out. The first is the compute: training and serving a 293B-A30B model entirely on domestic hardware means the whole pipeline runs on hardware available at home, not just the model weights ^[1]. The second is the shape of the release. A large hosted model, a per-token API price, and small open models for devices were announced together ^[1][2], which is a common way for a model family to cover data-centre, developer and local use at the same time.