SiliconFlow Models and Pricing Guide (2026): Free Tiers, Rate Limits & Cost Comparison

SiliconFlow's most compelling feature is its free model tier. These are not stripped-down demos — they are the actual production models, available at zero cost subject to rate limits. The free catalog rotates as new models are released, but as of mid-2026 these three are reliably free. The full SiliconFlow overview provides additional context on the platform's overall architecture.

Qwen3-8B (model ID: Qwen/Qwen3-8B) is the best all-rounder for free-tier users. It handles general conversation, text summarization, and light coding tasks well. DeepSeek-R1-Distill-Qwen-7B (deepseek-ai/DeepSeek-R1-Distill-Qwen-7B) provides chain-of-thought reasoning, making it ideal for math, logic, and step-by-step problem solving. DeepSeek-OCR (deepseek-ai/DeepSeek-OCR) specializes in extracting text from images and documents.

Rate limits for free models are relatively generous at the L1 verification tier — typically enough for personal development and testing. For heavier workloads, you either wait for the limit to reset or switch to a low-cost paid model.

When free models are not enough, SiliconFlow offers the latest flagship models from each major Chinese AI lab. Here is the pricing breakdown for the most popular ones (prices in CNY per million tokens, checked July 2026):

DeepSeek-V4-Pro (deepseek-ai/DeepSeek-V4-Pro) is the flagship reasoning and coding model from DeepSeek. At roughly ¥4 input / ¥16 output per million tokens, it delivers top-tier performance at a fraction of the cost of Western alternatives like GPT-4o or Claude Opus. This is the go-to model for production applications that need the best quality.

Qwen3-235B (Qwen/Qwen3-235B) is Alibaba's largest open-source model, excelling at multilingual tasks and long-context reasoning. GLM-5 (Zhipu/GLM-5) from Zhipu is strong for Chinese-language understanding and tool-use scenarios. Kimi K2.6 (moonshot/Kimi-K2.6) offers an exceptionally long context window for document-heavy applications. MiniMax-M2.5 (MiniMax/MiniMax-M2.5) is optimized for creative writing and dialogue.

SiliconFlow assigns every account a spending tier that determines API rate limits. Higher tiers unlock more requests per minute and higher token throughput. The tier is based on your cumulative spending and upgrades automatically as you spend more.

L0 is the default for new accounts with under ¥50 cumulative spending. L1 activates after basic verification. L2 through L5 correspond to spending thresholds of ¥200, ¥1,000, ¥3,000, and ¥10,000 respectively. Free models are subject to these limits regardless of whether you are paying, so heavy free-tier users may want to reach L2+ for more headroom.

For a concrete example: an L0 account might be limited to 1,000 requests per hour on a free model, while an L3 account gets 10,000+. Exact numbers vary by model and are shown in the console dashboard.

For prototyping and learning: start with the free Qwen3-8B. It covers 80% of use cases at zero cost and lets you build and test your application end-to-end before spending anything. If you don't have a SiliconFlow API key yet, the registration tutorial takes about five minutes.

For production coding and reasoning: DeepSeek-V4-Pro is the best balance of quality and cost. Its chain-of-thought reasoning and large parameter count make it competitive with models costing 5–10× more.

For Chinese-heavy applications: GLM-5 is purpose-built for Chinese language and cultural context, often outperforming Western models on domestic content.

For long documents (books, legal contracts, research papers): Kimi K2.6's extended context window handles inputs that would overflow other models.

For cost-minimized creative writing: MiniMax-M2.5 at ~¥1/¥1 per million tokens is hard to beat for volume text generation.

The most effective cost strategy on SiliconFlow is a tiered model approach: use free models for development and testing, then switch to paid models only for production traffic. This can reduce your development-phase API costs to literally zero. You can also explore using your key in other tools to test models in different environments without additional setup.

Second, monitor your token usage closely. Chinese-language prompts tend to use more tokens per character than English due to tokenization differences. If your application processes a lot of Chinese text, factor this into your cost estimates.

Third, take advantage of caching where possible. SiliconFlow does not currently offer prompt caching discounts like some Western providers, but you can implement client-side caching for repeated queries to reduce API calls.