MiniMax

public

china · Founded 2021-12 · IPO 2026 (HKEX: 0100) · Market cap: $39B (HKG: 0100) as of 2026-03

MiniMax is one of China's "AI Tigers," founded in December 2021 by a team of former SenseTime employees. CEO Yan Junjie (formerly SenseTime VP of Technology), COO Yeyi Yun (formerly SenseTime strategy), co-founder Zhou Yucong (Beihang University, formerly SenseTime algorithms R&D), and co-founder Yang Bin (formerly Uber ATG autonomous trucks). The company maintains only ~385 employees with an average age of 29.

MiniMax went public on the Hong Kong Stock Exchange (HKEX: 0100) on January 9, 2026, raising ~$620M with shares surging 109% on debut. Pre-IPO investors included Tencent, Alibaba (~12.5% stake), HongShan (Sequoia China), and miHoYo (Genshin Impact). Total pre-IPO funding was ~$1.15B.

MiniMax is distinctive for its "all-modality" approach and "Every FLOP Counts" efficiency philosophy. Its Lightning Attention architecture (7 linear + 1 softmax attention layers per block) reduces complexity from quadratic to linear, enabling context windows up to 4 million tokens. Over 70% of revenue comes from overseas through Talkie, an AI companion app with 236M+ users globally. The model series has progressed from ABAB-5 through MiniMax-M2.7, and Hailuo video generation has produced 600M+ creator videos.

Website Wikipedia HuggingFace GitHub Artificial Analysis
frontieropen-weightmultimodalreasoningefficiency

People

News

Outputs (19)

The MiniMax-M2 Series: Technical Report

paper

"Mini Activations Unleashing Max Real-World Intelligence." Comprehensive technical report covering the full M2 / M2.1 / M2.5 / M2.7 series — an…

MiniMax-M2.7

model

Current flagship "Co-worker Agent" focused on AI self-evolution. Can independently build and stabilize agent systems. 56.22% on SWE-Pro. Proprietary (…

MaxClaw

library

Integrated agent interface that natively calls MiniMax's video, audio, and image models.

MiniMax-M2.5

model

RL-trained on 200K+ environments, achieving 100 tok/s and 80.2% on SWE-Bench Verified. SOTA in office productivity (Word/Excel automation) and browse-…

↓ 490.6k

MiniMax-M2.1

model

Enhanced for multi-language programming (Rust, Java, Go) and complex agentic scaffolding. Open weights (229B total).

↓ 46.8k

VIBE & OctoCodingBench

dataset

Evaluation benchmarks for measuring agentic and coding performance.

Towards Scalable Pre-training of Visual Tokenizers

paper

Focused on the VTP (Visual Tokenizer) backbone for high-fidelity generation.

MiniMax-M2

model

First "agent-native" model from MiniMax (230B total, 10B active MoE). Designed for tool-calling and code execution, released at a fraction of the pric…

↓ 123k

MiniMax-M1 2

model paper

World's first large-scale hybrid-attention reasoning model (456B total, 45.9B active). Uses Lightning Attention for 1M-token contexts with 75% less co…

↓ 12.1k

SynLogic 2

paper dataset

Research on generating massive-scale logical reasoning data for training, with publicly available verifiable reasoning trajectories.

↓ 496

One RL to See Them All: Visual Triple Unified RL

paper

Introduces the V-Triune system and Orsta models (7B/32B) that unify visual reasoning and perception tasks via reinforcement learning. Up to +14.1 impr…

MiniMax-Speech: Intrinsic Zero-Shot TTS

paper

Research on learnable speaker encoders for high-fidelity voice cloning.

MiniMax Speech Series

model

Hyper-realistic vocal models supporting 40+ languages with instantly responsive latencies (<250ms). Versions from Speech 02 to 2.8.

Image-01

model

Text-to-image model designed for cinematic-quality visuals at extremely low cost.

MiniMax-01 3

model paper

A 456B parameter MoE model series featuring a 4-million-token context window and the Lightning Attention architecture. This work is highly influential…

↓ 101.3k 📄 1

Hailuo AI (Video-01 / 2.3)

model

Industry-leading video generation series. Hailuo 2.3 (Oct 2025) achieved breakthroughs in physical realism and facial micro-expressions.

MiniMax Music Series

model

Capable of generating full 5-minute studio-grade tracks with paragraph-level precision control. Versions from Music 1.5 to 2.5+.

ABAB 6 / 6.5

model

Transition to Mixture-of-Experts architecture, significantly boosting reasoning and math capabilities.

ABAB 5 / 5.5

model

Early foundational models that powered MiniMax's first consumer apps.