Physix FrontierPhysix Frontier · News Briefing Card (TMTPost · Nov 8, 2026)
iFlytek launches Spark X2.6, runs ultra-long context training on domestic chips
KEY FACTS
- iFlytek unveiled its device-cloud integrated Spark X2.6 at the 1024 Developer Festival, marking the first time million-token ultra-long context training has run on domestic chips.
- Spark X2.6 improves ultra-long context training efficiency by roughly 4x, and long chain-of-thought reinforcement learning efficiency rose from 30% to 84%.
- iFlytek's Q3 2026 revenue was 6.564 billion yuan, up 7.99% year over year, with net profit attributable to shareholders of 226 million yuan.
- Q3 non-GAAP net loss was 126 million yuan, and the cumulative non-GAAP loss for the first three quarters reached 763 million yuan.
- iFlytek acknowledged a roughly fivefold gap versus Nvidia's H200 in ultra-long context training above 256K.
KEY DATA
6.564 billion yuanQ3 revenue
226 million yuanQ3 net profit attributable to shareholders
126 million yuanQ3 non-GAAP net loss
roughly 4xUltra-long context training efficiency gain
PHYSIX OBSERVATION
The technical report card is impressive, but the commercial one is still up in the air. Domestic compute running million-token context is a hard breakthrough, yet persistent non-GAAP losses and a shrinking market cap show the market has shifted from listening to stories to counting revenue. iFlytek's project-based model is misaligned with token-based valuation logic, and an API-driven turnaround is unlikely in the short term; the key question is whether industry deployments can convert R&D spending into sustainable profit.
Source: TMTPost report