Model library
Explore models from our own compute and official channels, with current prices. Use one API key for every model.
25 models · Own compute 3 · Official channels 22
MiniMax H3 video generation supports text-to-video, image-to-video (first frame or first/last frames), and reference-to-video using text, images, audio and video. Output: 768P, with audio generated by default. Billed by successfully generated duration at a list price of CNY 0.50/second; promotional pricing is shown on the page. Failed generations are not billed. Reference images, audio and video inputs are temporarily free.
MiniMax H3: text/image/reference-to-video (first or first/last frames), with text/image/audio/video inputs. 768P; audio on by default. List CNY 0.50/s; promo price on page. Failures unbilled; reference inputs temporarily free. Fast: pull-mode queue; submit returns task id. Only 1344x768 (16:9, default) or 768x1344 (9:16); 5–14s (default 5s). Frame rounding may add up to ~0.7s; billed by requested seconds. Other parameters match H3.
GLM-5.3 is the next-generation model in Z.AI's GLM series, supporting up to 1M context tokens and 128K output tokens. Native deep thinking, tool calling and structured output support long-document understanding, code generation and complex reasoning. Served on our own compute infrastructure, with both OpenAI and Anthropic API protocols.
HappyHorse 1.1 image-to-video model creates a video from one image and a text prompt. Uses asynchronous task-based generation. Billed by successfully generated video duration: CNY 0.36/second at 720P or CNY 0.48/second at 1080P. Failed generations are not billed.
HappyHorse 1.1 reference-to-video model creates a new video from reference images or videos and a text prompt. Uses asynchronous task-based generation. Billed by successfully generated video duration: CNY 0.36/second at 720P or CNY 0.48/second at 1080P. Failed generations are not billed.
HappyHorse 1.1 text-to-video model with asynchronous task-based generation (submit → poll → download). Billed by successfully generated video duration: CNY 0.36/second at 720P or CNY 0.48/second at 1080P. Failed generations are not billed.
Qwen text-to-image model. Billed per successfully generated image at CNY 0.2/image; failed generations are not billed.
Qwen speech recognition model. Accepts audio via base64 (≤10MB) or a public URL. Transcribed text output is not billed.
Qwen text-to-speech model. Billed by input characters at CNY 0.1 per 1,000 characters. Supports multiple voices via the voice parameter; default: Cherry.
Qwen's visual understanding model with image input and thinking mode.
Qwen's omni-modal understanding model accepts audio, images and video; streaming is required. Output uses multimodal pricing; text-only calls use the same rates. Audio input is converted at the official rate of 7 tokens/second and billed separately.
A lightweight, low-cost Qwen model with near-flagship performance and a one-million-token context window.
Qwen's flagship model with its strongest reasoning capabilities and a one-million-token context window.
A balanced Qwen model with thinking mode and a one-million-token context window.
Qwen3.7 text embedding model converts text to dense vectors for retrieval, clustering and semantic matching. Billed by input characters at CNY 0.25 per million characters.
Qwen3.8 flagship chat model: a 2.4-trillion-parameter MoE with major improvements in coding and office tasks, native visual understanding, and deep semantic analysis of very long documents and videos. It can autonomously orchestrate multistep tasks and keep evolving. Supports deep thinking, tool calling, structured output and web search.
Tongyi Lab's text-embedding-v4 model supports multiple languages, including Chinese and English, with selectable vector dimensions (2048/1024/768/512). Suitable for semantic search, RAG, clustering and classification.
Wan2.6 text-to-video model with asynchronous task-based generation (submit → poll → download). Billed by successfully generated video duration: CNY 0.6/second at 720P, with 5/10-second options. Failed generations are not billed.
Wan2.7 image-to-video model creates a video from one image and a text prompt. Asynchronous task-based generation (submit → poll → download). Billed by successfully generated video duration: CNY 0.3/second at 720P or CNY 0.5/second at 1080P. Failed generations are not billed.
Wan2.7 image generation model creates images from text descriptions. Billed at CNY 0.10 per successfully generated image.
Wan2.7 reference-to-video model creates a new video from reference images or videos and a text prompt. Uses asynchronous task-based generation. Billed by successfully generated video duration: CNY 0.3/second at 720P or CNY 0.5/second at 1080P. Failed generations are not billed.
Wan2.7 text-to-video model with asynchronous task-based generation (submit → poll → download). Billed by successfully generated video duration: CNY 0.3/second at 720P or CNY 0.5/second at 1080P. Failed generations are not billed.
Wan3.0 all-in-one video generation model: text-to-video, image-to-video (first frame or first/last frames) and reference-to-video, with text, images, video and audio accepted together as references. Up to 30 seconds at 480P/720P/1080P; audio is generated by default (same price with audio on or off). Billed by successfully generated video duration; failed generations are not billed.
Wan3.0 high-speed video generation model. Matches the standard model's text-to-video, image-to-video and reference-to-video capabilities, accepting text, images, video and audio together, with significantly faster end-to-end generation. Up to 30 seconds at 480P/720P/1080P, with audio generated by default. Billed by successfully generated video duration; failed generations are not billed.
GLM-5.2 supports a larger context window and output length, with up to 1M context and 128K output tokens. It adds streaming output for tool calls via `tool_stream=true`, enabling real-time access to tool call arguments. Like the GLM-4.7 series, it supports deep thinking with `thinking={ type: "enabled" }`, where thinking is enforced once enabled, and introduces `reasoning_effort` to control reasoning intensity under chain-of-thought mode. It also delivers stronger coding performance and more advanced reason