SiliconFlow
Listed at https://siliconflow.cn
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | ¥0.4-2/1M tokens (Qwen2.5-7B ~¥0.4，Llama 3.3 70B ~¥2) | per million tokens |
| Output | ¥0.4-2/1M tokens | per million tokens |
| Cache Read | 部分模型支持 | Discounted |
🤖 Supported Models (100)
✨ Pros
- ✓All-in-one model hosting in China, 100+ open-source models online
- ✓Full Qwen3.5 / DeepSeek / GLM-4 / Llama 3 coverage
- ✓Direct access from China, no proxy needed
- ✓First-launch of Nex-N2-Pro (397B MoE reasoning model)
- ✓Pricing from ¥0.4/1M tokens, 80% lower than OpenAI official
- ✓OpenAI-compatible API
⚠️ Cons
- ×Overseas access requires proxy
- ×Does not host GPT/Claude series
- ×Model updates lag official by 1-2 weeks
- ×No first-class Function Calling support
- ×Pricing sensitive to FX fluctuations
🎯 Best For
China-based developers; teams wanting one-stop access to Chinese open-source models (Qwen3.5/DeepSeek/GLM); cost-efficient RAG/reasoning scenarios
💰 Pricing & Plans
| Model | Input (¥/M) | Output (¥/M) | Cache Hit | Notes |
|---|---|---|---|---|
| DeepSeek-V4-Flash | 1.00 | 2.00 | 0.02 | Flagship MoE, 1M ctx (est.) |
| DeepSeek-V4-Pro | 12.00 | 24.00 | 1.00 | Premium reserved capacity |
| DeepSeek-V3.2 | 4.00 | 6.00 | 0.40 | Prior-gen strong all-rounder |
| GLM-5.2 | 8.00 | 28.00 | 2.00 | Zhipu frontier coding model |
| Kimi-K2.7-Code | 6.50 | 27.00 | 1.30 | Moonshot code-specialist |
| Qwen3.5-122B-A10B | 0.80 | 6.40 | — | Cost-efficient MoE workhorse |
| Nex-N2-Pro | 1.75 | 7.00 | 0.17 | 397B MoE reasoning model |
| MiniMax-M2.5 | 2.10 | 8.40 | 0.21 | Balanced general model |
🔧 API & Developer Experience
- •OpenAI-Compatible API: A single endpoint mirrors the OpenAI chat-completions format, so most OpenAI SDKs work after swapping the base URL (https://api.siliconflow.cn/v1) and API key — minimal code change to adopt.
- •Unified Model Catalog: Over 100 open-source models (DeepSeek, Qwen, GLM, Kimi, Llama, MiniMax and more) behind one key and one billing account; no per-model signups or separate credentials.
- •Free Tier & Free Models: New accounts receive free credits on signup, and a curated set of serverless models is offered at ¥0 (bge embeddings, GLM-Z1-9B, OCR) for zero-cost prototyping and embedding workloads.
- •Pro Tier: A paid 'Pro' reservation tier buys higher TPQ/throughput and prioritized first-token latency for production traffic, addressing the throughput caps on the free shared pool.
- •Serverless, Pay-Per-Token: You never manage GPUs — SiliconFlow hosts 100+ models on shared clusters and bills per million tokens (input/output/cache), with auto-scaling for traffic bursts.
- •Fine-Tuning & Embeddings: Supported fine-tuning and a full embedding suite (bge-m3, rerankers) make it a one-stop shop for RAG pipelines, not just chat completions.
- •Realtime & Multimodal: Live realtime speech and vision-capable models are listed alongside text, so streaming/voice and RAG use cases share the same API surface.
🧠 China-Native Open-Source Model Hosting
SiliconFlow's core strength is being a China-native, no-infrastructure playground for the open-source ecosystem: it loads 100+ models — DeepSeek, Qwen, GLM, Kimi, Llama, MiniMax and more — onto shared GPU clusters and sells them per token over one OpenAI-compatible API. That removes the two big frictions Chinese teams face elsewhere: signing up for and managing rented GPUs, and cross-border latency. New accounts get free credits and a set of truly ¥0 models (bge embeddings, small GLMs, OCR) for zero-cost prototyping. The trade-off of this breadth is depth: because it mirrors many model families rather than betting on one, per-model optimization and first-token performance trail a dedicated first-party provider, and the fastest 'Pro' throughput requires a paid reservation tier.
🌐 China Access & Latency
SiliconFlow is a mainland-China-native service: the api.siliconflow.cn endpoint is directly reachable from within China without a proxy, with no GFW workaround and no cross-border egress charges, and latency to its Alibaba Cloud-hosted clusters is typically single-digit to low-teens milliseconds RTT. Billing is in RMB via prepaid recharge, and the same account serves the entire open-source catalog. For teams outside China, access follows the typical reverse direction — the mainland-hosted API generally needs a proxy from overseas, and the English documentation is thinner than the Chinese one, so non-Chinese teams may find setup and support lag. The platform is best treated as a China-first serving layer on top of the open-source model ecosystem rather than a globally distributed API.