SiliconFlow

SiliconFlow gives developers serverless access to open-weight LLMs, image, video and speech models through one OpenAI-compatible API, billed by usage.

SiliconFlow runs open-weight AI models on demand, so a developer can call DeepSeek, Qwen or Kimi without renting or managing GPUs. The company of the same name builds and operates it, and sells it as one platform under one signup. That platform has three parts: serverless inference, fine-tuning and reserved GPUs, the last listed as coming soon. Every model sits behind a single API in OpenAI's request format, so one key and one client reach the whole catalog. Serverless inference is the core. Its chat catalog spans DeepSeek, Qwen, GLM, Kimi, MiniMax, gpt-oss, Gemma and Hunyuan, each listed with its context length. Vision-language models read images sent inside a chat message, and reasoning models take a thinking_budget setting that governs how much they reason before answering. Function calling lets a model hand back a structured call to the developer's own tools, while a separate embeddings endpoint serves retrieval. Beyond text, the platform generates images with FLUX and Z-Image, short videos with Wan, and speech with Fish-Speech, CosyVoice and IndexTTS, and it transcribes uploaded audio; fine-tuning takes a customer's dataset and deploys the tuned model on the same account. Developers and small product teams who want open models in production without running inference servers are the main audience. Startups weighing several model families also fit, since moving between them means changing the model name in a request. Code already written against OpenAI's SDKs can move across with little rework. Buying is self-serve and usage based: language models are charged by the token, media models by the image, the video or the byte of speech, and the bill comes after use. New accounts start with a small credit, monthly spending limits cap the bill, and larger volumes are agreed with the sales team. Coding assistants and agents are among the jobs it is built for: a chat model with function calling plans a step, calls a tool and reads the result. Retrieval-augmented search pairs the embeddings endpoint with a chat model, so answers draw on a company's own documents. Content teams turn a prompt or a still image into finished visuals with the image and video models, under the same key as their chat workloads. For a team that wants many open models on one bill and one OpenAI-style API, SiliconFlow is the inference layer behind the application.