SiliconFlow

SiliconFlow is the inference cloud out of China: an OpenAI- and Anthropic-compatible gateway over a multimodal catalog — chat, vision, embeddings, image, audio and video — whose draw is the Chinese open-weight fleet, with DeepSeek, GLM, Kimi and Qwen checkpoints priced per million tokens in renminbi. The numbers are the argument: DeepSeek V4-Pro at ¥12 input / ¥24 output, GLM-5.2 at ¥8 / ¥28, Kimi K2.7-Code at ¥6.50 / ¥27, small Qwen MoE models from ¥0.40 / ¥3.20, images at ¥0.10–0.30 each, text-to-speech at ¥0.05 per thousand characters and time-of-day discounts on some models. Standard caps sit at 500 requests and 2 million tokens per minute, and the vendor claims its own inference engine cuts latency by up to 70% while tripling throughput.

The judgement depends on where you are. Inside China this is a mainstream choice: one API over the models that matter locally, aggressive prices, and an escalation path from shared endpoints to reserved capacity, on-premises installation and a bring-your-own-compute mode that separates network, compute and storage for compliance. Outside China it becomes a routing decision with caveats — renminbi billing, mainland-hosted endpoints, cross-border latency, and a separate international portal at siliconflow.com. The performance claims are the vendor's and were not benchmarked here, and for compliance-sensitive workloads the gating question is whether the vendor's registrations satisfy your regulator, not what the price list says.