The demonstration lasts a few minutes. A teacher types a request in ordinary language; a model of a solenoid appears on the screen and turns when a hand passes over a camera. The room applauds.
Then the chairman steps up and says something that is not a boast. For much of the past year, he says, his company was largely absent from the market everyone else was fighting over — writing code, running multi-step tasks — because the chips it trains on could not hold a long enough context.
That sentence is the most informative thing released at the launch.
Key takeaways
- iFlytek's Spark X2.6 is a 293B-parameter mixture-of-experts model with 30B active per token, trained end to end on domestic accelerators (iFlytek, 22 October 2026).
- iFlytek says a 2-trillion-token, 256K-context training run on its Feixing cluster fell from 4.5 months to 1.2 months, roughly four times faster (China Business Journal, October 2026).
- The comparison is iFlytek's own earlier pipeline, not Nvidia hardware; no cross-vendor benchmark, chip count or cost per token was published.
- On the serving side, iFlytek reports inference on Feixing-2 chips at 4.5 times the pre-optimisation rate and a KV-cache hit rate up from 63% to 95.7% (company disclosure).
- The 4B on-device model open-sourced on 1 September 2026 has passed 2.56 million downloads, with agent coverage at 41.9% against an industry 12.8% (iFlytek).
What "all-domestic" actually covers
The stack matters more than the adjective. Spark X2.6 runs on Feixing (飞星), a cluster iFlytek built with Huawei on the latter's Ascend (昇腾) ecosystem and describes as China's first 10,000-card domestic compute base, sized for trillion-parameter training.
Running inference on domestic accelerators was already routine. Training was not. Inference calls on capacity that already exists; training builds it, and it runs for weeks, which is why interconnect failures and memory ceilings surface there first. iFlytek's claim is about the training half, and it is narrow: the model was trained on that stack, and the company says this is the first time its code and long-horizon agent capabilities were reached on it.
The launch disclosed neither card count, pre-training duration nor cost.
Why the model is shaped this way, and what the 4× sits against
Architecture is where hardware constraints become visible. A 293B model that activates 30B parameters per token keeps the arithmetic per token small while keeping total capacity large. That trade pays when memory bandwidth, not compute, is the scarce input: fewer weights load per token, and the model keeps its depth.
Context length is the harder problem. A million-token window — the stated limit — needs a key-value cache far larger than domestic cards can hold. iFlytek's answers were engineering ones: dynamic sparse attention, which computes over a selected subset of the context, and a training scheme that allocates parallelism and memory by sequence length instead of fixing both in advance.
The cited workload:
- 2 trillion tokens of training data at 256K sequence length.
- Run on the Feixing cluster, which the company also uses for serving.
- Duration cut from 4.5 months to 1.2 months — a factor of about 3.8, rounded to four.
That is the headline number: one workload, against the company's own earlier pipeline.
The chairman's own framing is the tell. Others needed three months for a comparable model version, he said; his team needed a year. A fourfold gain applied to a fourfold deficit returns you to the starting line, not ahead of it.
What would make the figure portable is missing: throughput per card, cluster size, and a run against an equivalent Nvidia system. iFlytek published none of these, and its performance language stays qualitative — the model is placed against the global first tier — with no named evaluation harness.
Read the right way, the disclosure is candid: a company that says its previous pipeline took a year for a job others did in three months has published its own handicap estimate.
Where the cost actually sits
Training is a capital event; serving is the recurring bill.
On Feixing-2 chips, iFlytek reports inference efficiency 4.5 times the pre-optimisation baseline, and a KV-cache hit rate that rose from 63% to 95.7%. The mechanism is a four-tier cache that sorts context across memory types by access frequency, keeping hot context close and pushing cold context outward.
The hit rate matters because of how agent workloads are billed: each turn re-sends the accumulated context, and a cached prefix is not recomputed. Lifting the rate from 63% to 95.7% shrinks the recomputed share of every call, which is why iFlytek frames it as cost, not speed.
The same logic explains the on-device push. Spark X2.6-4B is open-sourced, and iFlytek says routing accuracy between phone and cloud rose from 81.2% to 96.1%, with on-device agent coverage at 41.9% against a stated industry baseline of 12.8%. A task finished locally is a cloud call not made.
What this means outside China
Three readings.
- For API buyers: the domestic-training claim is someone else's procurement problem. Serving price, cache behaviour and context quality are yours.
- For chip analysts: the useful disclosure is the deficit, not the fix. A gap that had to be engineered away says something measurable about the domestic stack on long-context work.
- For anyone modelling Chinese compute: sparse attention, tiered caching and on-device routing all reduce the same quantity — memory bandwidth per useful token. Expect more of this while memory stays scarce.
Honest limitations
Every efficiency figure here is company-disclosed at a launch and not independently replicated. The 4.5-to-1.2-month result describes one internal workload, not a standard benchmark, against iFlytek's own earlier pipeline. No chip count, total training cost or per-token price was published, and the performance claims name no evaluation harness. The download figure counts downloads, not deployments.
What readers can do now
- Ask any vendor claiming a domestic-stack speed-up for four numbers: tokens trained, sequence length, cluster size and wall-clock time — and whether the comparison is with their own prior pipeline or another vendor's hardware.
- When budgeting agent workloads, request KV-cache hit rates alongside list prices. At a partial hit rate, much of every call is recomputed and paid for twice.
- Test long-context claims on your own documents at the full stated window, because a context limit is a capacity statement, not a quality one.