MiniMax reset market expectations in the space of a week. The company released its multimodal generation model H3 on July 31, then said on Aug. 3 that it would open-source the model. By 10:53 a.m. on Aug. 7, MiniMax (00100.HK) was up 23.15% on the day. From the H3 launch through that point, the stock had risen 78.21%.
The moves were staggered across several sessions. Shares rose 13.15% on July 31, the day H3 was released. After the weekend, MiniMax announced on Aug. 3 that H3 would be open-sourced, and the stock added 7.20% that day. On Aug. 6, it jumped another 17.10%. The source article says the market has already begun to assign the company a new valuation framework.
Why H3 mattered more than another model ranking update
The article says that by July, investors had already started to reduce the weight given to single-model SOTA narratives in valuing domestic open-source large language model companies. Rankings were changing too quickly, and leaders were being overtaken in short order. Since Kimi K3, both DeepSeek and Qwen had released new models, yet valuation and market-cap swings were less dramatic than before.
MiniMax’s latest move landed differently because it centered on video generation. Rather than discounting the importance of a single model release, the market started to price video models into the valuation story.
One reason was performance. According to MiniMax’s website, H3 supports text, image, video and audio multimodal context, and can generate videos up to 15 seconds long at 2K resolution, 24FPS, with native stereo audio. In Artificial Analysis blind testing, H3 posted a 1242 Elo score in text-to-video with audio, ranking second globally. It ranked first in video editing and placed in the top three for image-to-video.
Price was another part of the explanation. The source article says H3’s per-second cost at 2K resolution is less than one-third of flagship models on the market, while its 768P pricing is less than half that of mainstream models.
Still, the article treats open-sourcing as the more consequential development. Once MiniMax released the weights on Aug. 3, chip vendors, inference frameworks, developer communities and enterprise service providers moved to support it. In the article’s telling, that lifted not only the company’s ceiling but the ceiling of the surrounding ecosystem as well.
Video generation and the hunt for the next high-token workload
The source places H3 inside one of 2026’s clearest AI market narratives: the push into coding. It describes the capital market’s reaction to that theme as benchmarking against Anthropic.
The article argues that the popularity of products such as OpenClaw and Claude Code reflects a basic shift in token usage. Coding tasks consume far more tokens. As model capability improves, demand for harder tasks rises and those tasks can be handled more effectively. Token usage per task rises too, and that feeds revenue growth. A Ping An Securities research note cited in the article says weekly global token calls were about 6.4T at the start of 2026 and had climbed to a weekly peak of 62.8T by July 2026, a 10x increase in half a year. Coding and agents accounted for the largest share of that consumption, and an agent task could consume 30 to 100 times as many tokens as an ordinary chat task. The article adds that Anthropic’s rapid ARR growth helped investors accept the logic of high-value tasks leading to high token consumption and then to high ARR.
Once that framework was established, investors began looking for the next dense token-consumption scenario. Video generation emerged as one candidate. Compared with text, video requires far more computation by nature. A Guotai Haitong research note cited by the article says a 1080P, 25-frame video can consume 30,000 to 50,000 tokens per second, and that figure could rise to 50,000 to 100,000 in the future. Seedance 2.0, according to the same article, consumes roughly 310,000 tokens to generate a single 15-second video in real-world testing.
That matters if video generation moves from novelty into production workflows in advertising, e-commerce, short drama, gaming and product presentation. The article says that in such a shift, video could become a new center of token consumption. It also points to a more direct path to monetization. Advertisers, e-commerce merchants, MCNs, game companies and brands already have content budgets. If AI video can reduce the cost of shooting, editing and asset production, customers do not need to wait for long enterprise IT budgeting cycles to start paying.
This is the backdrop for the rerating in MiniMax shares, according to the source. H3 enters a category that could scale token consumption quickly. If AI coding has already shown a path to commercializing text and code tasks, video generation may show a path to commercializing multimodal content production.
Open-sourcing H3 and the “DeepSeek moment” comparison
The article says that before H3, video generation models were largely controlled by closed-source vendors, including domestic players such as Seedance and Kuaishou Kling, as well as leading U.S. models. Closed-source systems help control the pace of monetization, but they also limit ecosystem spread. App companies depend on API pricing and usage rules, enterprise clients struggle to customize deeply, and cloud and chip providers can only adapt within narrow bounds.
That is why the article describes H3 open-source as something that changed the industry’s “historical process.” MiniMax said 16 chip vendors had completed adaptation work, including Huawei Ascend, Hygon, Moore Threads, MetaX, Kunlunxin, Biren, Iluvatar CoreX, AMD and Intel. It also said developer communities and cloud inference platforms such as HuggingFace, ModelScope, ComfyUI, RunningHub and fal had integrated H3, while inference frameworks including vLLM-Omni and SGLang offered support as well. More than 100 enterprises were live on Day 0.
That changed H3 from a model distributed only through MiniMax’s own API into one that developers, cloud providers, chipmakers and enterprise service firms could distribute together. The article says the effect closely matches the impact of the “DeepSeek moment” on the AI ecosystem in early 2025.
It also revisits an earlier market shift. According to the article, secondary-market investors became willing to compare domestic AI infrastructure hardware valuations with U.S. peers after DeepSeek V4 showed it could be trained on domestic chips. At the current stage, the source argues that mainstream domestic open-source large language models, with their advantages in cost-performance and ecosystem reach, have already threatened the position of the top two North American frontier model companies. In its view, that should push application-layer penetration higher.
The same change showed up in valuations. When the DeepSeek V4 preview was released in April, the market initially treated it as a negative shock to model-company valuations and shares in other model companies fell that day, seeing it as a competitor. By late July, when V4 Flash moved into formal release, the market had shifted to an ecosystem framework and started to treat it as a cost benchmark and capability floor for China’s open-source AI stack. The article argues that H3 follows a similar path, but this time in video generation.
Performance and cost remain central to that case. The source says H3 has entered the first tier of video generation models while undercutting mainstream alternatives on price. MiniMax says its 2K per-second price is less than one-third of mainstream models. A Guotai Haitong research note cited in the article gives a figure of about $0.13 per second.
The article also points to architectural efficiency. According to MiniMax’s official WeChat account, H3-VAE saw major efficiency gains. It says the first-generation H3 fully overhauled the previous tokenizer technology, improving both reconstruction and learnability. That gave H3 competitive efficiency, while a high compression ratio delivered a 4x sequence-length gain and materially reduced training and inference costs. The article says this is the key technical foundation behind native 2K output.
Open-sourcing on top of that does more than increase model calls. It changes strategic position. The article says closed-source commercialization in video has been difficult. If a model company relies only on API charging, it has to carry customer acquisition, compute, client service and industry adaptation costs itself. Open-source changes that division of labor. Cloud providers can host and serve inference, enterprise software firms can build vertical solutions, chipmakers can optimize domestic compute adaptation, and developers can build plugins and workflows.
In that setup, the article says MiniMax can exchange open-source access for ecosystem position, then monetize through enterprise customization, commercial licensing, model services and enterprise accounts. That path may not show up in financial statements immediately, but it can alter how the market classifies the company. The source ties this to the “Jevons paradox” narrative that became popular in early 2025: when token costs fall sharply, usage expands rather than contracts.
Sell-side reports shift toward revenue and ARR logic
Sell-side institutions moved quickly after the H3 launch. The article says Citi maintained a Buy rating on MiniMax in a July 31 report and sharply lifted its target price to HK$533, implying about 131% upside from the share price at that time. It describes Citi’s method as applying a 12x 2028E P/S multiple and forecasting a 128% revenue CAGR from 2025 to 2030.
AlphaEngine data cited in the article shows that multiple institutions issued positive ratings after H3 was released. Taken together with Citi and other reports, the article says the core post-H3 valuation logic still centers on model-driven revenue growth, specifically ARR expansion.
The source also looks for comparables in domestic closed-source video models. A Huaan Securities note cited in the article says models such as Seedance and Kling have reached a combined ARR on the order of nearly $3 billion. Based on that comparison, the article argues that if H3 maintains leading performance rankings and cost-performance advantages, MiniMax’s revenue under H3 could plausibly reach at least the same scale of growth. Under that growth rate, it says, MiniMax’s current P/ARR appears more attractive.
A longer-dated option: video models and physical AI
The article closes with a longer-term possibility. Video models, it says, may eventually become part of the foundation for physical AI.
Chris Paxton, a researcher at Agility Robotics, wrote on X: “The new Minimax H3 open video generation model looks like it has great potential for robotics. A lot of progress in robotics is driven by broader improvements in AI, and the current interest in world models means video generation can drive progress quickly.”
The article then adds a limit to that thesis. H3 supports text, image, video and audio, but compared with systems such as NVIDIA Cosmos that are aimed at world models and embodied intelligence training, video generation models like H3 still lack motion trajectories, first-person-view data and robot interaction data. For now, the “world model training foundation” story can only be treated as a long-dated option.
In the nearer term, the source says the market has already recognized the core value of H3. In its framing, the model does not just mark a turning point for a video-model-driven ecosystem. It may also mark the starting point of the next phase in MiniMax’s market-cap growth.