MiniMax M3
MiniMax's open-weight flagship LLM (稀宇科技), distinct from the Hailuo video product: 428B mixture-of-experts, 23B active per token, 1M context and native image and video input.
Overview
Hailuo is the consumer video product most people outside China know MiniMax by. M3 is a separate artifact: the lab's open-weight text and multimodal foundation model, released 1 June 2026 with weights on Hugging Face under the MiniMax Community License. Architecturally it is a 428B mixture-of-experts with roughly 23B parameters active per token, a 1M token context served through MiniMax Sparse Attention and native image and video input in the same checkpoint.
The reported numbers are strong for an open-weight release: 80.5% on SWE-bench Verified, 59% on SWE-bench Pro, 78.1% on MMMU Pro and 85.4% on Video-MME v2. MiniMax attributes a 9x prefill and 15x decode speedup over M2 to the sparse attention design. Self-hosting is a data-centre exercise rather than a workstation one. FP8 weights land near 230GB and the MXFP8 build near 440GB, so two H200s in FP8 will serve at 256K to 300K context while genuine 1M context serving wants four H200s or eight H100s with an FP8 KV cache.
Two practicalities matter if you are working with MiniMax from either side of the firewall. The console is regionally split, with platform.minimax.io and api.minimax.io for international accounts and platform.minimaxi.com and api.minimaxi.com for mainland ones. Keys do not cross over, and a mainland key pointed at the global host returns a bare invalid-key error rather than anything diagnostic. The second point is licence drift. M2 shipped MIT in October 2025 and M2.5 kept those terms in February 2026, then M2.7 landed in March 2026 with a licence requiring prior written authorisation for any commercial use at all. M3 sits between the two with its US$20 million threshold. Assume nothing carries forward.
Key Features
- ✓ 428B total parameters in a sparse MoE with roughly 23B active per token
- ✓ MiniMax Sparse Attention giving a 1M token context window, with a claimed 9x prefill and 15x decode speedup over M2
- ✓ Native multimodal input in one checkpoint: text, image and video
- ✓ Reported 80.5% on SWE-bench Verified, 59% on SWE-bench Pro, 78.1% on MMMU Pro and 85.4% on Video-MME v2
- ✓ Day-zero vLLM and SGLang recipes, plus lab-published FP8 (around 230GB) and MXFP8 (around 440GB) builds
- ✓ Regionally split consoles: platform.minimax.io with api.minimax.io for international accounts, platform.minimaxi.com with api.minimaxi.com for mainland
Where it holds
- • Weights are genuinely on Hugging Face rather than ModelScope only, with FP8 and MXFP8 variants published by the lab
- • Model card, licence text and deployment guides are written in English first, which is not the norm in this field
- • Frontier-class agentic coding scores at $0.30 per million input tokens on the hosted API
- • Two H200s in FP8 will serve it at 256K to 300K context, putting a 428B model inside a single-node budget
Where it breaks
- • The MiniMax Community License is not OSI open source. Free commercial use stops at US$20 million yearly revenue, and below that you still owe a one-time notice email
- • Commercial deployments must display the exact string 'Built with MiniMax M3' in a UI, website, about page or product docs. A logo alone does not count
- • Terms have tightened across the line: M2 shipped MIT in October 2025, M2.5 kept MIT in February 2026, then M2.7 arrived in March 2026 requiring prior written authorisation for any commercial use
- • True 1M context serving needs four H200s or eight H100s with an FP8 KV cache, which alone consumes roughly 120GB at full context
- • The licence bans military use categorically, which is a real procurement blocker for some defence-adjacent integrators
My Take
Open weights at this tier were not supposed to arrive this fast. 428B total with about 23B active, a real 1M token window and 80.5% on SWE-bench Verified, all pullable from Hugging Face in FP8. The catch is legal rather than technical: the MiniMax Community License caps free commercial use at US$20 million yearly revenue and requires a 'Built with MiniMax M3' credit in your interface. Read the LICENSE file on every release, because M2 was MIT and M2.7 was not.
Quick Info
- Pricing:
- freemium
- Openness:
- Open weights
- Licence:
- MiniMax Community License (custom, $20M revenue threshold plus attribution)
- Starting at:
- Weights are free to download and self-host. Commercial self-hosting is free below US$20 million yearly revenue after a one-time notice email to [email protected]; above that you need prior written authorisation. Hosted API is $0.30 per million input tokens and $1.20 per million output at up to 512K input, $0.60/$2.40 above 512K, $0.06 per million cache-read tokens, with a priority tier at 1.5x standard.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where MiniMax M3 holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to MiniMax M3
DeepSeek
freemiumChinese AI lab producing open-source LLMs that rival proprietary models at a fraction of the cost, with strong reasoning and coding abilities.
Kimi / Moonshot AI (月之暗面)
freemiumMoonshot AI’s long-context AI assistant, pioneering ultra-long document processing with a 2 million token context window.
Tongyi Qianwen / 通义千问 (Qwen)
freemiumAlibaba’s open-source large language model family and AI assistant, offering strong multilingual capabilities and deep enterprise integration with Alibaba Cloud.
Ling 3.0
open-sourceAnt Group's open-weight MoE family (百灵, Bailing): Ling-3.0-flash at 124B total and 5.1B active, a 7.9B tiny variant and trillion-parameter Ring reasoning models, all MIT.
Mistral AI
open-sourceEuropean lab publishing open weight models you can run on your own hardware. Large 3 is a 675B MoE with 41B active, 256K context and Apache 2.0 weights.