MiniMax

Overview

MiniMax is a leading AI assistant platform in China, created by the parent company of Hailuo AI, with strong capabilities in multimodal fields such as text, voice, and video. Its core model, MiniMax M3, adopts a new attention architecture MSA, supports 1M ultra-long context, and features native multimodal and cutting-edge coding capabilities, designed for complex tasks and collaborative scenarios. The platform also offers diverse tools such as video generation (Hailuo 2.3), voice (Speech 2.8), and music (Music 3.0), and meets different user needs through the API Token Plan.

Since 2026-07-31, MiniMax has expanded its generation stack with the H3 omni-modal video model, which was officially open-sourced on 2026-08-03. H3 jointly understands text, images, video, and audio, generating videos at up to 2K resolution, up to 15 seconds long, with 32 kHz native stereo audio, and it focuses on visual packaging and post-production effects such as dynamic MVs, dynamic posters, Vlogs, movie title sequences, and game UI. On 2026-08-27, the SGLang Diffusion team benchmarked MiniMax-H3 video generation on 8×NVIDIA H200, achieving a 1.85-1.95× speedup over Diffusers in the dense lossless path with no approximation loss. On 2026-08-31, a 24-hour AI livestream website powered by MiniMax H3 Max launched, with H3 Max 768P and 480P integrated into the open platform and MiniMax Design, and overseas developers have already used it to build Twitch live streams and 24-hour "AI TV stations."

Key Features

• MiniMax M3 Language Model: Adopts MSA sparse attention architecture, supports 1M ultra-long context, has cutting-edge coding and Agent capabilities, native multimodal mixed training, suitable for engineering-level collaborative tasks.
• MiniMax Code Coding Assistant: Desktop coding Agent that can autonomously form Agent teams, assign work based on task complexity, and remember user habits and preferences, enabling skill creation, memory viewing, and scheduled tasks within the dialog box.
• MiniMax H3 Omni-Modal Video Model: Officially released on 2026-07-31 and open-sourced on 2026-08-03, H3 is a general-purpose omni-modal generation system that jointly understands text, images, video, and audio, generating videos at up to 2K resolution, up to 15 seconds long, with 32 kHz native stereo audio. It focuses on visual packaging and post-production effects, capable of generating dynamic MVs, dynamic posters, Vlogs, movie title sequences, and game UI, with strong semantic adherence and strengths in instruction following, text and brand rendering, and V2V action transfer.
• MiniMax H3 Max Livestream Integration: On 2026-08-31, a 24-hour AI livestream website powered by MiniMax H3 Max launched. H3 Max 768P and 480P have been integrated into the open platform and MiniMax Design, and overseas developers have already used this to build Twitch live streams and 24-hour "AI TV stations."
• Hailuo 2.3 Video Generation: Latest video generation model, supports high-quality video content creation, suitable for creative and commercial scenarios.
• MiniMax Speech 2.8 Voice: Advanced speech synthesis model, provides natural and fluent voice output, supports multiple application scenarios.
• MiniMax Music 3.0 Music: Released on 2026-08-12, Music 3.0 is a new-generation, open-weight, production-grade all-around music model that can complete the composition, arrangement, performance, and production of an entire song in one go based on creative concepts and optional lyrics, supporting up to five minutes in length.
• API Token Plan Subscription: Offers flexible token subscription plans, Max version up to 7.1 billion tokens per month, priced at only 1/6 of similar products (such as Claude Max), supports cutting-edge models and ultra-long context.

Use Cases

• Developers use MiniMax Code for engineering-level code generation and collaboration
• Creators use Hailuo 2.3 to generate video content
• Creators use H3 to generate 2K video with native stereo audio for visual packaging and post-production effects
• Developers build 24-hour AI livestreams and "AI TV stations" with MiniMax H3 Max
• Enterprises integrate multimodal AI capabilities into products through the API Token Plan

Pros

• Comprehensive multimodal coverage, including text, voice, video, and music
• MiniMax M3 model supports 1M ultra-long context, suitable for complex tasks
• MiniMax Code features intelligent Agent teams and personalized memory functions
• API Token Plan offers high cost-effectiveness, Max version priced at only 1/6 of similar products
• H3 omni-modal video model is open-sourced and supports up to 2K resolution, 15-second videos, and 32 kHz native stereo audio
• H3 video generation achieves a 1.85-1.95× speedup over Diffusers on 8×NVIDIA H200 in the dense lossless path
• H3 Max 768P and 480P are integrated into the open platform and MiniMax Design for building 24-hour AI livestreams

Pricing

Free to use basic features; API Token Plan offers subscription, Max version ¥119/month, up to 7.1 billion tokens per month, supports cutting-edge models and ultra-long context. For H3 video generation, at 2K resolution the price per second is less than one-third that of mainstream models, and at 768p it is less than half the price of mainstream 720p offerings.

Summary

MiniMax is suitable for developers and creators who need multimodal AI capabilities, with core advantages in powerful coding Agent, ultra-long context support, cost-effective token subscription plans, and an expanding omni-modal generation lineup led by the open-sourced H3 video model.

Version History

• MiniMax M3.1 Flash released: nearly 3x faster inference at one third the cost (2026-10-19): MiniMax released M3.1 Flash on October 19, revealing it as the model previously tested anonymously on OpenRouter under the codename Space Bunny. Official figures put DeepSWE at 74%, Automation-Bench at 62%, ALE at 30% and BabyVision at 91%, with pricing of $0.15/$0.60 per million input/output tokens and $0.015 for cache reads — roughly three times the speed of M3 at one third the cost. During its anonymous run the model held first place on OpenRouter for 10 consecutive days and processed 56.7T tokens, and ranked first on OpenCode for 12 days with 106T tokens served. It shipped the same day on OpenRouter (free for 72 hours), OpenCode, Ollama, Cline and GMI Cloud, with GMI claiming day-zero support and a 70% reduction in agent task cost. Weights are scheduled for release on October 23, and a Pro version is in preparation.
• A 24-hour AI livestream website powered by MiniMax H3 Max has launched (2026-08-31): MiniMax has integrated H3 Max 768P and 480P into its open platform and MiniMax Design. Overseas developers have already used this to build Twitch live streams and 24-hour "AI TV stations."
• MiniMax-H3 benchmarked on 8×H200: 1.95× lossless speedup, up to 6.24× (SSIM 0.76-0.91 (2026-08-27): The SGLang Diffusion team benchmarked MiniMax-H3 video generation on 8×NVIDIA H200, achieving a 1.85-1.95× speedup over Diffusers in the dense lossless path, with no approximation loss.
• MiniMax Music 3.0 released: a new generation of open-weight, production-grade all-around music models (2026-08-12): MiniMax launches Music 3.0, a new-generation music generation model that can complete the composition, arrangement, performance, and production of an entire song in one go based on creative concepts and optional lyrics, supporting up to five minutes in length.
• MiniMax H3 officially open-sourced: a general-purpose omni-modal generation system supporting 2K video and native stereo audio (2026-08-03): MiniMax has officially open-sourced its new-generation general video model H3, which can uniformly understand text, images, video, and audio, generating videos at up to 2K resolution, up to 15 seconds long, with 32 kHz native stereo audio.
• MiniMax H3 released and set to go open source, focusing on visual packaging and post-production effects (2026-07-31): MiniMax releases AI video model H3 and announces open-sourcing. The model focuses on visual packaging and post-production effects, capable of generating dynamic MVs, dynamic posters, Vlogs, movie title sequences, game UI, etc., with extremely strong semantic adherence. In advertising and motion effects, it outperforms Seedance 2.0, but falls slightly short in cinematic tension shots. After H3 is open-sourced, a large number of specialized vertical-domain versions are expected to emerge.
• MiniMax H3 released: an open-source all-around multimodal generation model supporting 2K native stereo video (2026-07-31): MiniMax officially launches the all-in-one multimodal generation model H3, which can jointly understand text, images, video, and audio, generating videos up to 2K resolution, 15 seconds in length, with native stereo sound. H3 excels in instruction following, text and brand rendering, and V2V action transfer. At 2K resolution, its price per second is less than one-third that of mainstream models, and at 768p, it is less than half the price of mainstream 720p offerings. The company plans to open-source the model weights in the coming days to support the open-source community and accelerate hardware compatibility.