The China Academy of Information and Communications Technology (CAICT) has published its 2026 report on large-model token service performance. The report was produced with a Ministry of Industry and Information Technology key laboratory for large-model and hardware-software evaluation and with the Artificial Intelligence Industry Alliance. It draws on monitoring data from January to August 2026 and compares official model platforms with third-party providers serving four open-source models: DeepSeek, Qwen, GLM and Kimi.
Usage has grown quickly. According to CAICT, daily token consumption in China reached about 500 trillion in June 2026, nearly four times the level in December 2025. More model-as-a-service platforms now sell tokens through public cloud APIs, but the report says users still lack clear ways to judge service quality and often struggle to choose a provider.
Speed has improved. Average output throughput for domestic models reached 78.23 tokens per second in August, about double the 2025 average of 38.25. Latency has moved the other way: average time to first token was 1.26 seconds in August, up from 0.78 seconds in 2025. CAICT says this is not simply a decline in performance. Larger models, longer contexts and higher call volumes all pushed latency up at the same time, and the report found the same trend in typical models from other markets.
Prices have risen after two years of price cuts. CAICT says the normalized average price of domestic models was 6.8 yuan in the second half of 2025 and 13.46 yuan so far in 2026, an increase of almost 100 percent. Typical models from other markets launched in 2026 averaged about 117.8 yuan on the same normalized basis.
Context windows keep getting longer. Among the monitored models, 47.8 percent now support 1 million tokens and 95.7 percent support at least 200,000, more than 76 percentage points higher than at the end of 2025.
Performance also depends heavily on the platform. Throughput for the four open-source models kept rising after their May version updates. In August, DeepSeek-V4-Pro reached twice the throughput of DeepSeek-V3.2. Yet the same model's throughput, latency and success rate varied widely between platforms, with throughput gaps of nearly 2x, and some third-party platforms outperformed the official ones. CAICT attributes these gaps to differences in inference acceleration, optimization and cost control. All figures are from the CAICT report.
SEE ALSO: METASTONE Launches TokenFleet · SenseTime's SenseCore Details Prefill/Decode Split Clusters