MiniMax

What it is. MiniMax is the Shanghai-based Chinese general-AI company behind the M-series agent/chat models (MiniMax-M3, M2.7), plus video, audio, image and music models. Why it matters. MiniMax-M3 lists the cheapest flagship pricing in the database ($0.30/$1.20 per 1M tokens) while shipping open weights and 1M-token context — a cost-efficiency position distinct from both DeepSeek’s MIT-open stance and ByteDance’s closed premium. Key characteristics. The MiniMax Sparse Attention (MSA) architecture claims 9x prefill / 15x decode speedups at 1M context; open weights under the MiniMax Community License; a consumer-app portfolio (Talkie/星野). What a professional should know. The Community License is not MIT-style: commercial use requires attribution and, above revenue thresholds, written authorization — read it before commercial deployment.

MiniMax (上海稀宇科技有限公司), founded in early 2022, is a Chinese general-AI company whose products span chat/agent models (M-series), video (H-series), audio (Speech/TTS/ASR), image generation, music (Music 3) and consumer apps (Talkie / 星野). The company states it serves users in 230+ countries with 300M+ individual users and 2M+ enterprise clients and developers (vendor claim, CN site).

The current flagship model is MiniMax-M3 (2026-06-01): a ~428B/23B-active MoE with 1M-token context, multimodal input (text/image/video) and open weights under the MiniMax Community License. API platforms: platform.minimaxi.com (China, CNY) and platform.minimax.io (international, USD).

Why it matters

China AI Hub analysis: MiniMax’s structural role is the cost-efficient challenger that competes on both capability breadth and price without going full-open. It runs a broad product surface — MiniMax Code and MiniMax Agent on the agent side, plus video/audio/image generation — while keeping its flagship M3 at the lowest flagship price point in the database. China AI Hub analysis indicates MiniMax’s distinctive bet is architectural efficiency (MSA sparse attention) translating directly into price, which positions it as the value alternative to Alibaba Cloud’s integration breadth and ByteDance’s consumer premium. China AI Hub analysis indicates MiniMax’s Community License mirrors the same scale-conditioned openness as the other open-weight labs — attribution and, above revenue thresholds, written authorization keep it in the open ecosystem while bounding the openness at scale. MiniMax’s models are also served through the model-neutral platforms Dify and FastGPT, which can orchestrate M-series models alongside competitors’.

Entity hub

Models

• MiniMax-M3
• MiniMax-M2.7
• MiniMax-M2.7-Highspeed

Products

• MiniMax Code
• MiniMax Agent
• MiniMax open platform

API

• MiniMax API

Agents

• MiniMax Agent
• MiniMax Code

Research / Technology

• AI Chips
• AI Infrastructure
• Inference
• Mixture of Experts
• Multimodal AI
• Tool Calling

Comparisons

• Doubao Seed 2.1 Pro vs MiniMax M3
• Kimi K3 vs MiniMax M3

Pricing

• MiniMax pricing

Benchmarks

• BrowseComp
• SWE-bench
• Terminal-Bench

What is uncertain

• Funding-round details are not listed on official pages.
• The user counts (230+ countries, 300M+ users) are vendor claims stated on the CN site.

Sources

evidence_id	source_name	source_url	source_type	published	verified	confidence	conflict
src-companies-minimax-1	MiniMax official site (international)	https://www.minimax.io/	Official	—	2026-09-20	high	—
src-companies-minimax-2	MiniMax official site (China)	https://www.minimax.cn/about	Official	—	2026-09-20	high	—
src-companies-minimax-3	MiniMax API platform — model overview (CN)	https://platform.minimaxi.com/docs/guides/models-intro	Official documentation	—	2026-09-20	high	—
src-companies-minimax-4	MiniMax official release notes	https://platform.minimaxi.com/docs/release-notes/models.md	Official documentation	—	2026-09-20	high	—

Labels used above: Official fact (from MiniMax’s official sites and platform docs), Vendor-reported claim (the MSA speedup figures and user counts stated on the CN site), and China AI Hub analysis (our synthesis, always introduced as such).