The domestic LLM competition is unfolding on two battlefronts simultaneously: "chokepoint mitigation" and "edge deployment."

On September 7, iFlytek released the Spark X2.5 multilingual text generation model. Three key numbers: 293B-A30B MoE architecture, 256K context window support, coverage of over 200 languages, with focused improvements in coding and agent capabilities. The most critical line: end-to-end training and inference fully on domestic compute platforms.

Released the same day were two edge models open-sourced by subsidiary Ciyuan Xinghuo: Spark X2.5-4B and 1.7B, using a hybrid attention architecture, designed around agents, and the first edge models natively supporting up to 1 million tokens of context.

Flagship gets bigger, edge gets smaller, full stack domestic, open source. Four lines advancing at once — this isn't a single product launch, it's a complete positioning strategy.

Domestic Compute: From "Can Run" to "End-to-End Training"

For the past two years, the narrative around domestic LLMs has always been "fine-tuned on domestic compute" or "adapted to domestic chips" — it always sounded second-class. What X2.5 is saying this time is "end-to-end training and inference," meaning from pre-training through fine-tuning to production inference, not a single high-end overseas chip was used.

The significance of this isn't technical — it's an industrial signal. When domestic compute can actually support a 293B MoE-level model, it means three things:

First, the risk of domestic LLMs being chokepointed further decreases. No matter how much export controls escalate, there's at least a fallback path.

Second, the domestic chip ecosystem flywheel starts spinning. Model companies train models on domestic chips; the more they use them, the faster the chips improve; the faster the chips improve, the more model companies use them. In the past this flywheel was Nvidia + overseas model companies spinning it. Now the domestic version is starting to turn too.

Third, "self-reliance and controllability" goes from slogan to product spec. Going forward, for government, SOE, finance and other data-security-sensitive customers, "end-to-end domestic training" will become a hard requirement — not a bonus, a ticket to the table.

Edge Open Source: The Ambition of 1M Context

The most interesting point in this wave of edge model competition isn't "small yet powerful." It's "million-token context."

The historical pain point of edge models has been short context. With small parameter counts, they can't hold long context, so edge models can only do simple Q&A — they can't handle real agent tasks. Reading documents, processing long text, working with entire codebases — all of these require long context.

iFlytek this time pulled edge model context all the way to 1 million tokens, and it's native support, not compression, not a sliding window. What does this mean? It means edge agents finally have genuinely usable "memory."

Imagine: a 4B model running locally on your phone that can remember all your chat history, all your documents, all your calendar, then helps you with schedule management, email replies, information organization. All running locally, data never leaves the device, privacy never leaks. This is one of the ultimate forms of edge agents, and 1M context is the ticket to get in.

Of course, the overall capability of edge models is still far behind cloud flagships. A 4B model, no matter how optimized, can't beat a 293B MoE. But the battlefield of edge models was never about "who's stronger." It's about "who's closer." Strong models in the cloud handle complex tasks; small models on the edge handle personal贴身 tasks; the two are orchestrated through agent protocols in between — this is the hybrid architecture almost every big company is betting on. By open-sourcing million-token context on the edge side, iFlytek is essentially staking out a position in the "edge foundation" layer of this hybrid architecture.

Domestic LLM competition has already moved from "benchmark score wars" to "positioning wars." Some are competing on parameters, some on price, some on ecosystem, some on edge. iFlytek chose the "domestic + open-source edge" path — not the flashiest, but a very steady one. After all, in uncertain times, the words "self-reliant and controllable" are worth more than anything.

See you tomorrow.