SiliconFlow

SiliconFlow is a China-based inference API for open models. Rechecked 2026-08-17: the working official host is siliconflow.cn. siliconflow.com was not the page we used. The English models list at docs.siliconflow.cn/en/userguide/guides/about-models still sells an OpenAI-compatible endpoint. Example list prices that day: Qwen/Qwen3-8B ¥0.27 input / ¥0.88 output per 1M tokens, Qwen/Qwen3-32B ¥0.68 / ¥2.70. This page used to claim 6M users, 100B tokens/day, 2.3x faster, and 200+ models as if audited. Those figures are out.

Compare OpenAI if you want the US lab API, or Kimi if you want Moonshot's own stack.

Key Features

• OpenAI-shaped API: Drop-in base URL for many open chat models.
• Public CNY table: Per-model input / output on the docs page.
• Not a US hyperscaler: Latency and account rules follow a China host. Test it yourself.
• Model catalog moves: Do not freeze DeepSeek R-1-as-first-in-China as the 2026 story.

Use Cases

• Teams that need cheap open-model inference in CNY.
• Apps already on OpenAI SDKs that can swap the base URL.
• People who do not want to self-host vLLM.

Limitation: we did not re-verify funding rounds or Alibaba traffic-spike folklore. English docs and the .cn site can diverge. Recheck the model id you actually call.

Pricing

Examples from 2026-08-17 docs, per 1M tokens, CNY:

Recheck the live table. USD conversion is not first-party here.

Getting Started

1. Open siliconflow.cn and create a key.
2. Read the models guide.
3. Point an OpenAI-compatible client at their base URL.
4. Start with a cheap Qwen3 row before you copy a DeepSeek id from an old blog.

Frequently Asked Questions

Is siliconflow.com official?

Use siliconflow.cn until the company redirects the other host cleanly.

200+ models still right?

Do not quote a census. Open the current list.

Cheaper than official Qwen?

Sometimes on the rows we saw. Compare Alibaba Model Studio the same day.

Alternatives

• OpenAI: GPT-5.6 Sol API if you want that lab.
• Kimi: Moonshot consumer + API.
• Modal: If you wanted to host weights yourself.

Tips

1. Quote ¥/1M from the docs table, not "2.3x faster".
2. Keep monthlyVisits as the old catalog number, not a new MAU.
3. Recheck model ids. Qwen3-8B will not stay the cheap default forever.

Conclusion

SiliconFlow is a CNY open-model inference API with a public price list, not a 6M-user census. Start at siliconflow.cn and the models guide.