MiniMax’s AI Comeback: How H3 Turned a Post-IPO Selloff Into a Valuation Reset
A single product launch erased months of investor doubt: MiniMax, the Hong Kong-listed Chinese AI startup, has staged one of the most rapid stock recoveries in the city's nascent large-model sector, with shares surging 78.21% in the seven trading days following the July 31 release of its H3 multimodal generative model — reversing a prolonged slide that had seen the company's market capitalization shrink to within HK$9.5 billion of its first-day close.
The H3 launch on July 31 triggered an immediate 13.15% single-session gain, a response that stood in stark contrast to the market's brutal verdict on its predecessor. The rebound continued in subsequent sessions: shares added another 7.20% on August 3 when the company open-sourced H3's model weights, then jumped a further 17.10% on August 6 after MiniMax was formally included in the Hong Kong Stock Connect southbound trading scheme — a structural catalyst that broadens its mainland investor base. As of August 10 close, the stock traded at HK$322.40, giving the company a market capitalization of approximately HK$112.6 billion (US$15.6 billion).
Dissecting the Collapse That Made the Comeback Necessary
To understand the magnitude of MiniMax's reversal, investors must trace the arc from its January 9, 2026 listing — one day after peer ZhipuAI debuted — to its nadir seven months later.
MiniMax listed on the Hong Kong Stock Exchange to extraordinary fanfare. Shares surged as much as 109% on debut day, briefly pushing market capitalization above HK$100 billion, while ZhipuAI's stock nearly broke its issue price at launch. The market's initial read was unambiguous: MiniMax, with its consumer-facing global product suite and overseas revenue exceeding 70% of total sales, warranted a premium internet-style valuation multiple. At peak, MiniMax's market cap touched HK$410 billion (US$56.9 billion), more than double ZhipuAI's concurrent valuation.
That premium proved fragile. On March 2, MiniMax delivered what it billed as the world's first annual report from a large-model company. Shares gained 9.10% the following session. But founder, CEO, and CTO Yan Junjie's articulation of a "intelligence density × token throughput" platform thesis at the earnings call failed to map a credible monetization timeline. By March 4, shares had shed 10.48%, closing at HK$735 and erasing roughly HK$27 billion (US$3.75 billion) in market value in a single session.
The real pressure arrived in the June–July window. The June 1 release of the M3 model disappointed on multiple dimensions: while MiniMax claimed a 59% score on SWE-Bench Pro — purportedly surpassing GPT-5.5 — independent evaluator Artificial Analysis ranked M3 ninth among mainstream models on its Intelligence Index. More damaging was the company's pricing misstep: M3 launched at roughly double M2.7's price point, raising subscription costs from RMB 29/month to RMB 49/month. Within one week, MiniMax reversed course entirely, announcing a permanent 50% price cut and acknowledging "insufficient communication and inadequate transition planning."
Then came the lock-up expiry. On July 9, MiniMax faced its first major post-IPO share unlock, covering approximately 63% of Hong Kong-listed shares and 46.4% of total shares outstanding. The stock fell 17.98% that day. By comparison, ZhipuAI's first tranche of unlocked shares represented just 5.76% of its float — a fraction of the selling pressure MiniMax absorbed. By July 13, the valuation gap between the two companies had widened to a peak of 10.51 times, with ZhipuAI at HK$733.4 billion versus MiniMax at HK$69.8 billion. MiniMax hit its all-time post-IPO low of approximately HK$67.4 billion on July 20.
JPMorgan cut MiniMax from Overweight to Neutral in June, set a HK$400 target price, then lowered it further to HK$300 in July. Citigroup trimmed its target to HK$533 in the same month.
H3 Wins on Three Dimensions That Matter to Institutional Buyers
MiniMax's rehabilitation rests on a product that outperforms on the three variables institutional buyers weight most heavily: benchmark performance, unit economics, and ecosystem reach.
Performance: On the Artificial Analysis video model leaderboard — widely regarded as the most rigorous third-party benchmark — H3 achieved a video editing Elo score of 1,127, ranking first globally. Its text-to-video-with-audio score of 1,237 places it second only to Alphabet's Gemini Omni Flash at 1,243, while its image-to-video score of 1,193 trails only ByteDance's Dreamina Seedance 2.0 720p at 1,199. On the Design Arena platform, H3 ranked first among open-source models across three categories: multi-image-to-video, image-to-video, and video editing.
Notably, ByteDance released its own Seedance 2.5 model on the same day, July 31, setting up an immediate head-to-head comparison across the AI video creator community. In side-by-side tests using identical inputs, H3 drew favorable assessments in cinematic trailer generation, complex text rendering, and UI animation — use cases directly relevant to MiniMax's target verticals of advertising, e-commerce, and gaming.
Pricing: H3's video generation is priced at RMB 0.80 per second at 2K resolution — approximately one-third the cost of comparable flagship video models in the market. At 768P resolution, pricing is less than half the mainstream benchmark. Critically, MiniMax attributes the cost reduction to engineering innovation rather than subsidized pricing: its proprietary H3-VAE high-compression tokenizer achieves industry-leading video compression, reducing sequence length by a factor of four for equivalent video content. Combined with optimizations in heterogeneous training, load balancing, and GPU utilization efficiency, the architecture delivers margin-positive unit economics at the lower price point — a distinction analysts flagged as materially important.
Open-source adoption: Within 24 hours of H3's August 3 open-source release, over 100 domestic and international partners had deployed compatible integrations. Hardware coverage spans Huawei's Ascend chips, Moore Threads, AMD, and Intel. Developer ecosystem integrations include Hugging Face, ModelScope, and ComfyUI. On Hugging Face, H3 climbed to the top trending position, surpassing DeepSeek V4 Flash. Stable Diffusion co-founder Emad Mostaque publicly endorsed the release. The open-source move also positions H3 as what MiniMax describes as China's first open-source video generation model reaching the global first tier — a claim that, if sustained, carries meaningful geopolitical and commercial weight as Western restrictions on Chinese AI exports remain a live variable.
Institutions Revise Upward, but Execution Risk Remains the Central Question
The institutional reaction to H3 was swift. Jefferies reiterated a Buy rating with a HK$1,118 price target, arguing H3 solidifies MiniMax's competitive positioning in AI video generation and that its cost advantage will accelerate enterprise client acquisition and commercial deployment. Citigroup maintained its Buy/High Risk rating. Orient Securities published research projecting accelerated penetration of H3 into short-drama production, digital advertising, visual design, and visual effects.
The valuation gap with ZhipuAI has also compressed meaningfully. At MiniMax's August 10 close of HK$322.40, the market cap ratio between ZhipuAI (HK$582.96 billion) and MiniMax (HK$112.6 billion) stood at approximately 5.18 times — down sharply from the peak of 10.51 times recorded on July 13, though still a substantial premium for the B2G-oriented competitor.
The structural divergence between the two companies' business models remains unresolved. ZhipuAI's focus on enterprise and government private-deployment contracts — characterized by high per-client value, customized delivery, and localization — insulates it from consumer pricing volatility but constrains scalable margin expansion. MiniMax's consumer-and-API model offers higher operating leverage but demands continuous product iteration to sustain user retention and pricing power.
On July 10, Yan Junjie issued an internal letter titled "Toward the Edge of the Sky," pledging zero personal compensation until MiniMax achieves AGI, contributing 4% of his personal shareholding to an employee incentive pool, and establishing a 1% fund to support the open-source community. The same day, the company announced a HK$16 billion (US$2.22 billion) convertible bond financing round, with 80% of net proceeds earmarked for AI infrastructure and model R&D. The conversion price, set near the five-day average price prior to announcement, reflected limited pricing power — a reminder that investor confidence, even now, remains conditional.
One model launch can reset a week of expectations. It cannot, by itself, define a company's terminal value. MiniMax must now demonstrate that H3's performance benchmarks translate into durable revenue growth, that its open-source ecosystem generates defensible commercial relationships, and that its engineering team can sustain the iteration cadence required to remain competitive against ByteDance, Alphabet, and a domestic field that is compressing benchmark gaps at an accelerating pace.
Related Coverage:
MiniMax Races Toward 2.7 Trillion-Parameter Model as A-Share Listing Window Converge