GLM-5.3 strengthens Zhipu's capability advantage, while DeepSeek repricing provides MiniMax with short-term breathing room

Overview

The report evaluates the competitiveness of China's large-model companies through a Pareto frontier defined by capability and cost. Zhipu moves further toward the capability frontier with GLM-5.3, supporting the decision to maintain Overweight; although MiniMax benefits from DeepSeek repricing and the multimodal option represented by Hailuo H3, its M3 model has not yet established a clear advantage in capability or cost performance, so Neutral is maintained.

Core views

The report favors the capability frontier over a pure low-price strategy: frontier models can support higher pricing and unlock more complex workloads, while mature capabilities gradually become commoditized and face more providers and more intense price competition. Zhipu's improvement comes from internal model iteration and is more defensible; MiniMax's improvement partly results from changes in the external competitive environment, and its long-term investment case still depends on validation of M3.1 and paid monetization of Hailuo H3.

Analysis framework

The capability–cost Pareto framework is used, with mixed-token pricing as a measure of cost and benchmarks and observed product performance used to assess capability. The framework is intended for relative competitive positioning rather than precise measurement of customer return on investment.

Methodology notes

• Competitive analysis frameworkCapability–cost Pareto frontier

  A model lies on the frontier when competitors cannot provide higher capability at the same or lower price, or comparable capability at a lower price.

  Models on the frontier are more likely to have defensible commercialization potential; models inside the frontier can easily be squeezed by stronger or cheaper substitutes.

• Valuation methodologyP/E valuation and WACC discounting

  Based on expected 2030 P/E multiples and discounted back to the target date using a 15% WACC.

  Target prices for both companies are based on 20x expected 2030 P/E, reflecting the valuation premium assigned to high growth.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Z AI Co Ltd - H（2513.HK）

  Core beneficiary

  Strengths

  GLM-5.3 improves coding and agent capabilities; repeated releases of domestic frontier models establish a track record of ongoing iteration; its position on the capability frontier can create pricing and high-value workflow opportunities.

  Weaknesses

  GLM-5.3 performance and paid conversion still require independent validation; ongoing R&D investment and profitability pressure remain substantial.

  Comparison

  Compared with MiniMax, the latest improvement in competitiveness is driven more by internal model iteration; compared with frontier competitors such as Kimi and DeepSeek, continued updates are still necessary.

  Risks

  Export controls, geopolitics, and entity-list risks; intensifying competition; commercialization and customer adoption below expectations; dependence on computing power and external suppliers.

• MiniMax Group Inc - H（0100.HK）

  Watchlist name

  Strengths

  Covers text, audio, images, video, and productivity scenarios, with 2C and 2B product positioning, overseas expansion, and a multimodal business option; initial feedback on Hailuo H3 has been positive.

  Weaknesses

  M3 has not yet established a clear advantage in capability or cost performance; independent providers have weaker value-capture ability when facing integrated platforms such as ByteDance and Kuaishou.

  Comparison

  Compared with Zhipu, model capability remains in a catch-up phase; DeepSeek repricing eases cost pressure, but the advantage is not created by the company itself and may reverse.

  Risks

  M3.1 performance below expectations; insufficient paid multimodal use, pricing, or unit economics; litigation with U.S. film studios; competition, R&D investment, commercialization, and computing-power supply risks.

Key data

• Zhipu rating and target priceOverweight; HK$1,800Target price raised from HK$1,600; 2026–2030 revenue forecasts raised by 6%–10%.
• MiniMax rating and target priceNeutral; HK$260Target price raised from HK$160; 2027–2030 revenue forecasts raised by 11%–21%.
• Zhipu valuation referenceApproximately 20x US$5bn 2027E ARRThe report's implied ARR valuation reference.
• MiniMax valuation referenceApproximately 12x US$1bn year-end 2026E ARRThe report's implied ARR valuation reference.
• Current share price2513.HK: HK$1,270; 0100.HK: HK$329Both are prices as of 2026-08-14.

Impact & implications

Industry competition is shifting from individual model launches toward the ability to repeatedly sustain capability leadership or structural cost advantages. For Zhipu, post-training, data quality, reinforcement learning, evaluation systems, and engineering execution are becoming key to sustained differentiation; for MiniMax, competitor price increases alone are insufficient to create a long-term moat, and it must demonstrate that model iteration and multimodal commercialization can generate independent and sustainable profit pools.

Risks

• Model capability iterations may not receive independent validation or may lag competitors.
• API price competition may intensify again, eroding assumptions for retention, market share, and monetization.
• High R&D and computing-power investment may extend the loss-making period and suppress profitability.
• Enterprise customer adoption and paid conversion may fall short of expectations.
• Integrated internet platforms may leverage distribution, advertising, content ecosystems, and data advantages to compress the value capture of independent model companies.
• Export controls, geopolitics, supply-chain, and legal-litigation risks.

What to watch

• Independent evaluations, customer feedback, and paid conversion for GLM-5.3 in real-world coding and agent tasks.
• Whether Zhipu can continue launching frontier models and improve cost efficiency through inference optimization.
• Whether MiniMax M3.1's capability, API pricing, and relative cost performance are sufficient to enter or approach the Pareto frontier.
• Hailuo H3's sustained multimodal performance, paid usage, pricing durability, and unit economics.
• DeepSeek's subsequent pricing and model updates, and whether its structural inference-cost advantage persists.