MiniMax M3¶

Development line: project:minimax-m3 · thread minimax-m3
Last event: 2026-06-13 · 1 dated since 2026-06-13 · Researched: 2026-09-05 · confidence: medium

What it is¶

MiniMax M3 is an open-weight native-multimodal MoE model for developers building coding agents, document/repository analysis, and image or video-input workflows.

• Text, image, and video input.
• Hosted Anthropic-compatible and OpenAI-compatible APIs.
• Local serving through SGLang, vLLM, Transformers, KTransformers, or Unsloth.

1M-token context; about 428B total parameters and 23B activated parameters; the official Hugging Face checkpoint is BF16 under the minimax-community licence.

Start with the hosted API; self-host only after sizing hardware and context capacity.

Development line¶

• 2026-06-13 — MiniMax M3 public project, model, and API resources appeared. On 2026-06-13, MiniMax M3 was linked through a public GitHub project repository, a Hugging Face model page, and MiniMax text-generation documentation. These links connect the project source, model distribution, and usage documentation surfaces.

What changed¶

• 2026-06-01 — MiniMax M3 launched in MiniMax Code, Token Plan, and the API; MiniMax said the weights would follow within ten days.
• 2026-06-13 — the public GitHub repository, Hugging Face checkpoint, and API documentation were listed together; the upstream GitHub history shows no separately dated model or documentation change on this date.
• 2026-06-18 — KTransformers was added to the documented local-deployment choices.
• 2026-06-22 — an Unsloth tutorial was added to the documented local-deployment choices.

How to use this¶

As of 2026-06-13, evaluate MiniMax M3 against its linked repository, Hugging Face model page, and text-generation documentation before deciding whether to adopt it.

1. Create a MiniMax API or subscription key and select MiniMax-M3 as the model. — https://platform.minimax.io/docs/guides/text-generation
2. For the recommended API path, send an Anthropic Messages request to https://api.minimax.io/anthropic/v1/messages; use the OpenAI-compatible /v1 endpoint only when your client requires it. — https://platform.minimax.io/docs/guides/text-generation
3. Set thinking to adaptive for mixed workloads, enabled for deliberate reasoning, or disabled when latency and throughput matter more. — https://github.com/MiniMax-AI/MiniMax-M3
4. For local inference, download MiniMaxAI/MiniMax-M3 and serve it with a supported runtime such as SGLang or vLLM. — https://huggingface.co/MiniMaxAI/MiniMax-M3

Best practices¶

• Begin with the official sampling defaults: temperature=1.0 and top_p=0.95. — https://github.com/MiniMax-AI/MiniMax-M3
• Treat the 1M context window as a deployment target, not an automatic setting: choose accelerator, dtype, parallelism, cache, and maximum context for the actual traffic shape. — https://github.com/vllm-project/vllm-project.github.io/blob/main/_posts/2026-06-12-minimax-m3-vllm.md
• Probe provider-specific tool support in the exact client before depending on it; an open issue reports that Anthropic-style built-in web search was unsupported despite Messages API compatibility. — https://github.com/MiniMax-AI/MiniMax-M3/issues/23

Superseded by this¶

• 2026-06-01 — guidance that an M2-series model is the latest M-series default is superseded for new M3-targeted long-context, multimodal, coding, and agentic work. Current documentation keeps M2 models available rather than declaring them retired.

Still unknown¶

• No first-party release note or upstream commit identifies a distinct MiniMax M3 model, weights, or API change on 2026-06-13; we treat the date as an access-link step, not a launch date.
• The 2026-07-01 and 2026-07-11 README commits have generic messages, so their user-visible changes cannot be stated reliably.
• MiniMax has not published a single current first-party hardware-sizing table for the official M3 BF16 checkpoint across target context lengths.

Sources¶

Agent brief¶

• Subject: project:minimax-m3, thread minimax-m3, 1 dated events 2026-06-13 → 2026-06-13.
• Practical note: As of 2026-06-13, practitioners should evaluate MiniMax M3 against its linked repository, Hugging Face model page, and text-generation documentation before deciding whether to adopt it.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.