Huachuang Securities: MiniMax H3 Model Open Source, Industry Competition Moves Towards Productivity

Huachuang Securities published a research report stating that MiniMax has officially opened the H3 base model weights. The key to H3 is not merely improving resolution, but rather through the synergy of models, data, and infrastructure, transforming complex full-modal generation into a low-cost product that can be scaled. The competition in video models is shifting from single-instance material generation to understanding complete creative intent and participating in production processes in areas such as advertising, e-commerce, gaming, and design. Domestic large models have entered a new phase where reasoning efficiency and open-source ecology are equally emphasized. Industry competition has transitioned from simply enhancing intelligence limits to a comprehensive comparison of model stratification, full-modal generation, agent execution capabilities, and the cost of individual tasks. The main points from Huachuang Securities are as follows: Unity of tasks and modalities: video models are transitioning from generation tools to universal creative systems. H3 can uniformly understand the multi-modal context composed of text, images, videos, and audio, generating videos of up to 15 seconds long with a maximum resolution of 2K and native stereo sound. Unlike traditional models that separate text-to-video, image-to-video, action references, audio references, and video editing into multiple independent tasks, H3 allows users to directly describe the relationships between materials using natural language, completing subject retention, action transfer, audio reference, and content editing within the same model. The report suggests that the competition in video models is shifting from single-instance material generation to understanding complete creative intent and engaging in production processes in advertising, e-commerce, gaming, and design. Comprehensive architecture for task generalization, with efficiency advantages forming an important foundation for H3's commercialization. H3 establishes relationships between different materials and target content through Contextual Omni Representation, employing a 33B parameter single-stream H3-Omni Transformer for joint prediction of video and audio; H3-VAE's high compression ratio offers a 4-fold benefit in sequence length, and the heterogeneous training architecture for understanding and generating allows for nearly a 30% increase in end-to-end training throughput. In the 2K output stage, H3 reintroduces 768p results into the original context for re-generation to enhance the recovery capability of small text and fine details. The report emphasizes that the key to H3 is not just improving resolution but rather through the collaboration of models, data, and infrastructure, transforming complex full-modal generation into a low-cost product ready for scaling. Opening the base model accelerates ecological diffusion; MiniMax is exploring the synergy between open source and commercialization. This opening is centered on H3-Base and supports text-to-audio-video generation, beginning and ending frame generation, and full-modal reference generation, compatible with mainstream deployment frameworks like SGLang, vLLM, Diffusers, and ComfyUI. The H3-Context-IR responsible for complex input orchestration and the H3-Regenerate-2K responsible for 2K re-generation are currently provided via official APIs. Within 24 hours after the opening, nine chip manufacturers, including Huawei Ascend, Moore Threads, Muxi, Haiguang, AMD, and Intel, completed Day 0 adaptations, with over 100 developer communities, cloud inference platforms, and enterprises having completed access. The report posits that "opening the base model while retaining high-value orchestration and HD services" can both expand ecological influence and reserve a clear commercialization entrance for API calls and enterprise services. In the past two weeks, new models have launched in force, with performance upgrades and price reductions accelerating in tandem. On July 30, OpenAI reduced the price of GPT-5.6 Luna by 80% and Terra by 20%, after adjustments, Luna is priced at $0.20 per million input tokens and $1.20 per million output tokens; Terra's API price is $2 per million input tokens and $12 per million output tokens; Sol's price remains unchanged. On July 24, Anthropic launched Claude Opus 5, supporting 1M context and 128K outputs, priced at $5/$25, emphasizing strong enhancements in complex agentic coding and enterprise tasks. Domestically, Qwen launched the 2.4T parameter, 1M context Qwen3.8-Max on August 3; DeepSeek-V4-Flash-0731 went live on August 1, targeting high concurrency and low-cost scenarios with 284B total parameters, 13B active parameters, and 1M context. The report suggests that industry competition is transitioning from simply enhancing intelligence limits to a comprehensive comparison of model stratification, full-modal generation, agent execution capabilities, and unit task costs. Risk Warning: Technological progress may not meet expectations; model implementation may fall short of expectations; business implementations may not meet expectations.