MiniMax released its first open-source multimodal generative model, H3, on July 31. This model supports processing text, images, audio, and video, offering direct output in 2K resolution and generating audiovisual content up to 15 seconds long. The release of H3 followed the M3 model, which was presented five months earlier, but its unsatisfactory performance caused a 15% drop in stock value in a single day.
In July, the industry faced a certain 'line of survival': Moonshot AI unveiled the weights of the Kimi K3 model, while DeepSeek lowered the cost of the V4-Flash cache to 0.02 yuan per million tokens. The Kimi K3 model, boasting 2.8 trillion parameters, became the world's first open model at the trillion-parameter scale, ranking fourth globally and first among open-source models. The output cost of the DeepSeek V4-Flash model is about $0.28 per million tokens, which is roughly 90 times cheaper than Claude Opus 4.8. A model that cannot surpass Kimi K3 in performance and lacks DeepSeek's pricing strategy risks disappearing from the market.
Model H3 occupies an intermediate position: it ranks first globally in video editing according to the Artificial Analysis leaderboard, enabling 2K generation for 0.8 yuan per second, which is three times cheaper than comparable flagship solutions. Nevertheless, it is not a leader in either overall performance or price. The choice of open source is viewed as a more passive step aimed at lowering thresholds in the AI field, attracting chip manufacturers and developers for optimization compatible with domestic chips, and securing its place in the video segment dominated by Seedance through ecosystem and developer scale. The commercial value of open source will be tested by financial indicators.
Open source is a double-edged sword: in the long term, it creates ecosystem advantages deeper than the patents of closed systems, as demonstrated by Linux and Android. However, in the short term, it foregoes direct monetization, and open source itself does not generate revenue. Open-source companies typically have revenues an order of magnitude smaller than their closed competitors: OpenAI earns about $13 billion, and Anthropic about $45 billion annually, while Seedance 2.0's monthly revenue exceeds 100 million yuan. Open source can gain reputation, community, and developers, but so far has failed to ensure profitability.
MiniMax's 2025 financial reports show strain: revenue reached $79.04 million, a 158.9% increase, with gross profit at $20.08 million and a margin of 25.4%. However, Research and Development (R&D) expenses rose by 33.8% to $253 million, mainly due to cloud training costs and more than triple the volume of output. Adjusted net loss in 2025 reached $251 million compared to $240 million in 2024: revenue increased by one and a half times, while losses remained at the same level, driven by investments of similar scale rather than efficiency gains. Building an open-source platform requires developer tools, community operations, corporate support, and chip adaptation, and each aspect demands capital. The question arises as to how long shareholders will patiently wait for a company forced to continue burning cash in search of a more valuable direction. MiniMax has bet on overcoming the line of survival through ecosystem and extreme cost-performance ratio, rather than parameter scale.
MiniMax has released the weights of the open-source H3-Base model on the HuggingFace platform. This model is a dense, single-stream Omni-Transformer with 33 billion parameters, equipped with a VAE featuring 16x spatial and 4x temporal compression. Furthermore, its usage cost via API is one-third of the price of Seedance 2.0.
The H3 model is a universal system for generating all modalities, unlike models designed only for text-to-video or image-to-video conversion. It is capable of understanding contexts from text, images, video, and audio, and then generating video with native stereo sound. Architecturally, H3 does not consist of a sequential connection of video and audio models.
The main network, H3-Omni-Transformer, is implemented as a dense, single-stream Transformer with 33 billion parameters, with about 13 billion parameters located in the AdaLN branch, whose modulation results can be pre-calculated and saved for lighter inference. Text encoding is handled by H3-Encoder, visual data by H3-Encoder and H3-VisualVAE, and audio by H3-AudioVAE; all these elements are integrated into a single sequence that is jointly predicted for video and audio latents.
MiniMax improved H3-VisualVAE by implementing 16x spatial compression and 4x temporal compression. After additional patching, the effective spatial downsampling of video tokens entering the Transformer reaches a 32x level, which reduces computational costs during high-resolution video generation. The complete MiniMax H3 suite includes three components: H3-Context-IR, H3-Base, and H3-Regenerate-2K. H3-Context-IR interprets user-provided text, images, video, and audio, structuring complex multimodal instructions into intermediate representations suitable for generative models. H3-Base is responsible for the actual generation of video and audio with primary output at 768P resolution. The H3-Regenerate-2K component uses context regeneration to obtain a 2K version from the 768P result while preserving the original context. Currently, the H3-Base component has open weights.
The pricing policy makes H3 a significant market disruptor: 2K resolution generation costs 0.8 yuan per second, and 768P resolution is cheaper, with free audio input and up to five free images. In terms of performance, H3 demonstrates competitiveness against ByteDance's Seedance 2.0 at one-third the cost, and local deployment on two RTX 5090 cards or one RTX 6000 reduces costs to zero. With its flagship capabilities, favorable pricing, and open weights, the initiative for video creation becomes accessible to every developer.
The release of this model intensifies competition in the AI video generation market. Seedance 2.0 previously dominated due to unified joint audio and video generation, and MiniMax H3 directly challenges this position by offering comparable functionality at significantly lower costs. For content creators and developers, H3 provides flexibility unavailable on closed platforms, including the option for local deployment for data-sensitive workflows, custom fine-tuning, and integration with existing production pipelines. The open-weights strategy aligns with the global trend where Chinese AI companies use open source and aggressive pricing to counter closed ecosystems. Thus, H3 is reshaping the economics of AI video generation, challenging the assumption that advanced video models must be expensive cloud services, and potentially accelerating adoption in advertising, e-commerce, film production, and content creation, where usage was previously limited by cost sensitivity.
The investment strategy in the field of artificial intelligence is changing: the focus is shifting from celebrating increased efficiency to an arms race in computational power. The Kimi K3 model with 2.8 trillion parameters has become a turning point, comparable to the moment of DeepSeek 2.0.
After DeepSeek-R1 caused panic in the markets by suggesting that cheaper training would devalue investments in computing, the situation has changed. Jevons Paradox demonstrates that increased efficiency leads to an increase in overall consumption of computational resources, rather than a decrease in demand.
The volume of capital expenditures exceeds previous expectations. Google has increased its CapEx forecasts for 2026 by $15 billion, bringing them to the range of $195–$205 billion. Company management warns that infrastructure costs will continue to pressure margins due to rising depreciation and data center operating expenses, and they are expected to increase significantly in 2027.
Analysts predict that the total capital expenditures of the five largest global cloud service providers will exceed $800 billion in 2026, potentially reaching $1–$1.2 trillion by 2027. Meanwhile, the top 14 providers show an annual CapEx growth exceeding 90%, as the imbalance between supply and demand for AI computing power persists despite efficiency advancements.
The market has clearly divided between AI model developers and AI infrastructure providers. In June 2026, hyperscaler stocks fell by a cumulative 18%, marking the worst monthly performance since Meta's IPO in 2012, as investors questioned the possibility of proportional returns on massive capital investments. Even Google's earnings report, which exceeded expectations, triggered a 4% drop after the close of trading solely due to CapEx news.
Conversely, shares of AI hardware manufacturers rose by 16% during the same period, creating a mirror image of the trend. The most obvious winners were memory and semiconductor manufacturers, as the demand for HBM, DRAM, and GPUs remains structurally undersupplied.
The K3 model with 2.8 trillion parameters requires at least 10 GB300 GPUs for local deployment. Even if training costs decrease by 40%, doubling the training frequency within the arms race will lead to a 20% increase in net compute demand. Thus, efficiency stimulates more competition, not less.
The most acute bottleneck is memory: the supply of HBM from SK Hynix is limited for the next two to three years. AI inference is driving a resurgence in central processors, projected to grow 4–5 times. A Chinese example demonstrates the Zhipu AI data center with a capacity of 1 GW, showing that even companies focused on model efficiency are now actively investing in scaling computational power.
