MiniMax¶

Development line: organization:minimax · thread minimax-models-and-products
Last event: 2025-12-27 · 3 dated since 2025-06-18 · Researched: 2026-09-05 · confidence: medium

What it is¶

MiniMax provides AI models and services. Developers get an Anthropic-SDK-compatible API alongside video and speech services; creators get separate Hailuo and Audio products.

• M3: language, tool use, image/video input, and a 1M-token context window.
• H3: text, image, first/last-frame, and multimodal-reference video generation.
• Speech 2.8: synthesis, cloning, and voice design.

Development line¶

• 2025-06-18 — MiniMax surfaced MiniMax-M1 model resources. On 2025-06-18, MiniMax linked a Hugging Face collection for MiniMax-M1 and a GitHub repository for the model. The same dated link set pointed to Hailuo Create and MiniMax Agent, placing the model alongside video and agent surfaces.
• 2025-06-22 — MiniMax publicly surfaced its Audio product. Speech remains a distinct API line with synthesis, cloning, and voice design; music access has changed materially.
• 2025-12-27 — MiniMax surfaced M2.1 through hosted and local access routes. On 2025-12-27, MiniMax linked its MiniMax-M2.1 news page and Hugging Face model repository. The dated resource set also pointed to MiniMax Agent, an Anthropic-compatible text API reference, OpenRouter, and Ollama, listing multiple access routes for M2.1.

What changed¶

2024-10-10 — A MiniMax event has no retained text or URL, so it makes no source-backed historical claim. MiniMax's closest primary baseline is Video-01, released 2024-08-31 with text-to-video and image-to-video at 720p/25 fps for up to six seconds. This is context, not confirmation of the 10 October change.

2025-06-18 — MiniMax M1 and Hailuo 02 entered the record. M1 was an open-weight hybrid-attention reasoning model with 456B total parameters, 45.9B active parameters, and 1M context. Hailuo 02 added native 1080p video options.

2025-06-20 — Hailuo Video Agent entered beta with prebuilt video templates as Stage One. Later autonomous stages were a roadmap, not verified capabilities.

2025-06-22 — A MiniMax Audio endpoint appeared, but the page lists no dated release details. Speech remains a separate API line with synthesis, cloning, and voice design; music access has changed materially.

2025-12-27 — M2.1 entered the record as an open-weight, agent-oriented M2 model for multilingual coding, tool use, instruction following, and long-horizon tasks. MiniMax's release note is dated 2025-12-23, while its model release notes state 2025-12-22.

Found today — M3 released on 2026-06-01 as the current 1M-context language model. H3 open-sourced on 2026-08-03 as the multimodal video line. Music API access changed on 2026-08-20 for new users.

How to use this¶

As of 2025-12-27, evaluate MiniMax models by each listed route—MiniMax Agent, the text API, OpenRouter, or Ollama—rather than assuming one interface covers every product.

1. Choose the current surface first: M3 for language, coding, tool use, or multimodal API work; H3 for video; Speech 2.8 for voice. Do not start from historical M2.1 or Hailuo names. — https://platform.minimax.io/docs/guides/models-intro
2. Create a pay-as-you-go API key on the MiniMax Platform to access language, video, speech, and image models. — https://platform.minimax.io/docs/api-reference/api-overview
3. For language integrations, set ANTHROPIC_BASE_URL=https://api.minimax.io/anthropic in the Anthropic SDK, use the MiniMax API key, and call MiniMax-M3. — https://platform.minimax.io/docs/api-reference/text-anthropic-api
4. For tool calling, append the full assistant response—thinking, text, and tool-use blocks—to history before sending tool results and the next turn. — https://platform.minimax.io/docs/api-reference/text-anthropic-api
5. For H3 video, create an asynchronous task, save the task_id, query status, and fetch the output URL after success. Pick inputs and duration within the 4–15-second range. — https://platform.minimax.io/docs/api-reference/api-overview
6. For speech, use synchronous T2A for short interactive audio and asynchronous synthesis for long text. Do not start a new Music API integration; use MiniMax Audio or the open Music 3 model instead. — https://platform.minimax.io/docs/api-reference/api-overview

Best practices¶

• Treat Anthropic compatibility as a documented subset: the API ignores top_k, stop sequences, MCP servers, context management, and containers rather than emulating them. — https://platform.minimax.io/docs/api-reference/text-anthropic-api
• Use M3's adaptive thinking for complex reasoning and disable it for low latency; keep thinking blocks unchanged across tool-use turns. — https://platform.minimax.io/docs/api-reference/text-anthropic-api
• Use M3—not M2.x—for image or video input, and estimate tokens before sending large multimodal requests. — https://platform.minimax.io/docs/api-reference/text-anthropic-api
• Treat M2.1, Hailuo 02/2.3, and older Speech models as legacy, requiring an explicit compatibility or migration decision. — https://platform.minimax.io/docs/guides/models-intro
• Treat video generation as a job workflow instead of a synchronous call: keep task IDs, poll status, and handle cancellation or deletion cleanly. — https://platform.minimax.io/docs/api-reference/api-overview
• If a deployment requires local M2.1 weights, serve via SGLang or vLLM and review the Modified MIT license before shipping. — https://huggingface.co/MiniMaxAI/MiniMax-M2.1
• Check account eligibility before building music workflows: new users cannot access paid Music APIs, and free endpoints are discontinued. — https://platform.minimax.io/docs/api-reference/api-overview

Superseded by this¶

• 2024-08-31 / 2024-10-10 — Video-01 limits (720p, 25 fps, up to six seconds) are legacy. H3 covers multimodal inputs, 4–15-second outputs, 24 fps, and up to 2K.
• 2025-06-18 — M1 as default hosted 1M-context model is historical. MiniMax documents M3 as the current 1M-context language model; M1 is archival or self-hosted open weights.
• 2025-06-20 — Hailuo Video Agent Stage One beta and the Stage Two/Three roadmap are historical. Check current features against H3 and API docs rather than assuming the roadmap shipped.
• 2025-12-27 — Choosing M2.1 as a flagship model is obsolete. It remains API-compatible and downloadable, but MiniMax classifies it under Legacy Models.
• 2025-06-22 — Starting a Music API workflow from older MiniMax Audio guidance is obsolete: from 2026-08-20, paid Music APIs are closed to new users and free endpoints are discontinued.

Still unknown¶

• The 2024-10-10 event has no retained link or text. The 2024-08-31 Video-01 page is context, not proof of the October change.
• The links https://hailuoai.video/create, https://agent.minimax.io/, and https://www.minimax.io/audio returned JavaScript shells during retrieval, leaving current consumer UX and availability unverified.
• The 2025-06-20 X post was not readable. A same-day first-party Hailuo Video Agent announcement exists, but exact correspondence is unverified.
• The M2.1 release date varies in records: 2025-12-27 in the dated event, 2025-12-23 in the release article, and 2025-12-22 in API release notes.
• minimax and minimax_models mix corporate history with model lineages. They connect, but they do not form a single-product lifecycle.

Sources¶

Agent brief¶

• Subject: organization:minimax, thread minimax-models-and-products, 3 dated events 2025-06-18 → 2025-12-27.
• Practical note: As of 2025-12-27, practitioners should evaluate MiniMax model use by the explicitly listed route—MiniMax Agent, the text API, OpenRouter, or Ollama—rather than assuming one interface covers all MiniMax products.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.