SiliconFlow
What is SiliconFlow?
SiliconFlow is an AI Model Serving & Inference platform. Applications call hosted models over a REST API (https://api.siliconflow.com/v1) and receive generated tokens, embeddings, rankings, images, audio, or video. The same account also supplies AI Infrastructure & Development: managed LoRA fine-tuning jobs, reserved GPUs, and documented Bring Your Own Cloud (BYOC) isolation.
Key Capabilities
- Serverless inference: Pay-per-use calls with no cluster to provision. The OpenAI Python SDK works by setting base_url to https://api.siliconflow.com/v1 and a SiliconFlow API key. POST /chat/completions streams Server-Sent Events; reasoning models also emit reasoning_content. POST /messages exposes an Anthropic-style messages API. Bearer authentication. Documented error classes include 429 rate limit, 503 overload, and 504 timeout.
- Model coverage: Hosted families include DeepSeek (R1, V3 and later), Qwen (including Qwen2.5-Coder and Qwen3), GLM, Kimi, MiniMax, Hunyuan, FLUX and related image models, Wan video, CosyVoice2 and Fish Speech, plus embedding and rerank models such as Qwen3-Embedding. The cloud Models page is the live catalog; the API enum can change as models are added or retired.
- Chat features: JSON object mode (response_format: json_object), function calling (tools with OpenAI-compatible function schemas), prefix completion, and Fill-In-the-Middle (FIM) via prefix/suffix on /completions or extra_body on /chat/completions. Reasoning models accept thinking_budget for chain-of-thought length. DeepSeek-R1 max_tokens can be set up to 16,384; docs recommend not setting max_tokens to the full context window.
- Retrieval endpoints: POST /embeddings returns vectors for strings or token arrays (example models: Qwen3-Embedding-8B/4B/0.6B). POST /rerank ranks a document list against a query.
- Image generation: POST /images/generations for FLUX.2 Pro/Flex, FLUX.1 Kontext, FLUX 1.1 Pro/Ultra, FLUX.1 schnell/dev, Qwen-Image, and Z-Image. Returned image URLs expire after one hour.
- Video generation: POST submit returns a request ID; the client polls a status endpoint. The generated result is available for 10 minutes; the download URL is valid for one hour.
- Speech: POST /audio/speech (text-to-speech) with speed 0.25–4.0, gain −10 to 10 dB, and formats mp3, opus, wav, pcm. Eight system voices (alex, benjamin, charles, david, anna, bella, claire, diana) are selected as model:voice. Users can upload a reference clip (recommended 8–10 seconds, under 30 seconds for CosyVoice2) as base64 or file. POST /audio/transcriptions transcribes uploaded audio.
- Fine-tuning: UI or API jobs for chat models (Qwen2.5-7B/14B/32B/72B-Instruct at documentation time) and image-generation models. Training data is .jsonl with alternating user/assistant messages. Configurable learning rate, epochs (1–10), batch size (1–32), max tokens, and LoRA rank/alpha/dropout. Validation can be a 10% split or a separate set. The resulting model ID is called on /chat/completions.
- Reserved GPUs: Dedicated, always-on GPU capacity for long-running or high-volume work, billed separately from serverless tokens. Isolated dedicated endpoints are listed on the products page as coming soon.
- TypeSafe / System One (Alpha): POST /systemone runs the Kev-4B decision model on mixed yes/no (noul), choice, and score questions. The TypeSafe SDK is documented as free through 2026-10-08 unless pricing is announced later.
- Playground and keys: Cloud console at https://cloud.siliconflow.com/ for model list, online test, playground (language, text-to-image, image-to-image), API key creation, and fine-tune jobs.
Audience & Use Cases
- Audience: Application developers and platform teams that need a hosted inference API, plus teams that fine-tune Qwen (or image) checkpoints and deploy them on the same endpoint.
- Use Case: Chat and reasoning in production; code completion with FIM; RAG via embeddings plus rerank; agent tool calling; Dify, Claude Code, Cline, DB-GPT, OpenClaw, and similar clients pointed at the SiliconFlow base URL; image and video generation APIs; TTS with cloned voices; LoRA fine-tunes that shrink long prompts.
Technical Specifications
- Base URL: https://api.siliconflow.com/v1. Auth: Authorization: Bearer from https://cloud.siliconflow.com/account/ak.
- SDKs: OpenAI Python library (Python 3.7.1+). TypeSafe Python SDK for System One (TYPESAFE_API_KEY or a SiliconFlow key).
- Billing: Serverless token prices (input, cached input, output per million tokens) and per-image prices on the pricing page. The vendor states $1 in starting credits and no minimum spend on on-demand usage.
- Deployment options in docs: Serverless public API; reserved GPUs; BYOC with compute, network, and storage isolation; hybrid cloud.
- Open-source engines (company, not the hosted SKU): OneDiff (diffusion inference) and BizyAir (LLM/multimodal runtime).
Categories & Use Cases
Technical Details
| Deployment Types | SaaS |
|---|---|
| Mobile Application | No |
FAQs
What is SiliconFlow?
SiliconFlow is an AI Model Serving & Inference platform. Applications call hosted models over a REST API (https://api.siliconflow.com/v1) and receive generated tokens, embeddings, rankings, images, audio, or video. The same account also supplies AI Infrastructure & Development: managed LoRA fine-tuning jobs, reserved GPUs, and documented Bring Your Own Cloud (BYOC) isolation.
What are SiliconFlow's top competitors?
Amazon Bedrock, Alibaba Cloud Model Studio, and DeepInfra are common alternatives for SiliconFlow.