MiniMax API for Agentic LLMs and Video

The MiniMax API is a single interface to MiniMax's model family, covering the M-series large language models for text reasoning and agentic coding and the Hailuo models for video generation. On Atlas Cloud you reach all of them through one OpenAI-compatible endpoint, so a single integration handles both text and video workloads. Pricing is pay-as-you-go per call, with no subscription or commitment.

The lineup splits into two tracks. On the language side sit the M-series models, including M2, M2.5, and the newer M3, built for coding agents, tool use, and long-context reasoning. On the video side are the Hailuo models, such as Hailuo 02 and Hailuo 2.3, which produce short cinematic and stylized clips.

Yes. Every MiniMax model on Atlas Cloud, from the M-series LLMs to Hailuo 02 and Hailuo 2.3, runs under the same API key and base URL. You can move between a text reasoning call and a video generation call inside one pipeline without setting up separate authentication.

Create an Atlas Cloud account, generate one API key, and point your existing OpenAI-compatible client at the Atlas base URL. Because the MiniMax API follows the OpenAI request format, most integrations need only a base URL and model name change rather than a rewrite. Start building today.

Hailuo 02 Pro generates native 1080p video and supports both 6-second and 10-second clip lengths per request, while the Standard tier trades resolution for lower cost. Hailuo 2.3 continues 1080p output with stronger handling of stylized and character-driven motion. Video is billed per second of generated footage, so cost scales with the length you actually produce.

MiniMax M3 introduces the MSA (MiniMax Sparse Attention) architecture, native multimodal input for images and video, and a context window MiniMax documents at up to 1M tokens. The M2 series, including M2.5, targets high-throughput coding and agentic tasks and is text-only. Choose M3 for long-context, multimodal agents and the M2 series for cost-efficient, high-volume coding pipelines.

Hailuo 02 leans cinematic, built on MiniMax's NCR rendering architecture for photorealistic, live-action style output. Hailuo 2.3 keeps that physics foundation but broadens into anime, illustration, and game-CG styles with finer control over character motion and micro-expressions. Reach for 02 on product and live-action work, and 2.3 on stylized or animated content.

Hailuo 2.3 earned the Physics Champion title on WorldModelBench for its accuracy in simulating mass conservation, fluid dynamics, and spatial-temporal consistency. Rather than approximating motion visually, it models physical behavior, which holds up in demanding scenes like tipping liquids, drifting objects, and complex body movement. For action-heavy or technically precise prompts, that physical fidelity is often the deciding factor.

MiniMax documents M3 at up to 1M tokens of context, enabled by its sparse attention design. The earlier M-series models expose a large context window in the low hundreds of thousands of tokens, though the exact configured limit can vary by provider. Check the specific model page on Atlas Cloud for the context length in effect for your deployment.