GLM-5.3 Post-Training Upgrade Strengthens Capabilities, but Neutral Maintained Amid Competition and Valuation Constraints

Overview

This report assesses the implications of Z.AI Co.'s GLM-5.3 release for model capabilities, productization, and valuation. Goldman Sachs believes this version achieves a performance step-up at an unchanged scale of 744 billion parameters through longer task environments, longer training duration, and reinforcement learning framework upgrades, demonstrating progress in post-training scaling for Chinese large models.

Core views

GLM-5.3 has improved on open-source coding benchmarks, long-horizon agent tasks, and cybersecurity capabilities, approaching global frontier levels at lower per-task costs. Goldman Sachs believes ZCode serves not only distribution and monetization functions, but also generates long-horizon, real-world coding trajectories from user interactions that can strengthen the model post-training data flywheel. At the industry level, improved model capabilities and API/subscription price adjustments indicate increasingly rational pricing. Although Zhipu offers model performance-to-price advantages, a data flywheel, and long-term commercialization potential, narrowing competitive differentiation, R&D investment, and cash burn constrain near-term risk-reward, supporting a Neutral rating.

Analysis framework

The report combines observations of model benchmarks and product functionality, industry pricing comparisons, peer ARR valuation multiple comparisons, and DCF valuation. The DCF uses a 12% WACC and 2% terminal growth rate, deriving the target price from market-share and long-term margin assumptions, while also presenting bull, base, and bear scenarios.

Methodology notes

• Valuation methodsDCF Valuation

  Determining a target price based on discounted future cash flows

  The HK$1,610 target price uses a 12% WACC and 2% terminal growth rate; key assumptions include 21% market share by 2030 and a 26% adjusted EBIT margin by 2035.

• Operating AnalysisData Flywheel

  Product usage data feeds back into model training

  ZCode usage can generate long-horizon, real-world coding trajectories expected to feed back into subsequent model post-training, creating a closed loop among product, data, and model capabilities.

• Relative ValuationP/ARR Comparison

  Measuring valuation by comparing market capitalization with annual recurring revenue

  The report compares Zhipu's approximately 30x expected year-end 2026 ARR with MiniMax's approximately 14x expected ARR, and believes Zhipu's valuation premium is partly supported by model capabilities and stage of development.

Asset mapping & comparison

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

• Z.AI Co. (2513.HK)

  Research Coverage

  Strengths

  GLM-5.3 post-training scaling improves coding, agent, and cybersecurity capabilities; the ZCode data flywheel strengthens model iteration; model performance-to-price advantages and long-term commercialization potential provide support.

  Weaknesses

  Near-term earnings visibility is limited, with high R&D spending and cash-burn pressure; the open-source model may cause disclosed ARR to understate actual demand, while also delaying revenue realization.

  Comparison

  The report compares Zhipu's approximately US$75 billion market capitalization and approximately 30x expected year-end 2026 ARR with DeepSeek's approximately US$50 billion most recent valuation and MiniMax's approximately US$14 billion market capitalization and approximately 14x expected ARR.

  Risks

  Global foundation-model competition, narrowing performance gaps with peers, weaker-than-expected commercialization, financing and cash-burn pressure, and geopolitical risks related to China-U.S. technology competition.

Key data

• Investment RatingNeutralRating maintained.
• 12-Month Target PriceHK$1,610Based on DCF valuation.
• Valuation ParametersWACC 12%; terminal growth rate 2%Consistent with Goldman Sachs coverage methodology.
• GLM-5.3 Parameter Scale744 billion parametersUnchanged from GLM-5.2; the performance improvement is attributed to post-training scaling.
• ZCode Active UsersMore than 1 millionThe report describes it as a coding tool product optimized for GLM.
• GLM Coding Plan Lite PriceRMB118/monthCredit-based subscription; based on the company's indicated weekly token allowance, it is approximately 50% cheaper than equivalent API costs.
• Implied Bull/Bear Case ValuationHK$3,266 / HK$835Scenario analysis based on key assumptions.

Impact & implications

Improved model capabilities, better task cost efficiency, and the product data flywheel are expected to support Zhipu's competitive position in coding and enterprise use cases and may improve long-term monetization capacity. If industry price increases persist, they would benefit model vendors' revenue quality and profitability trajectory; however, narrowing model performance gaps mean valuation premiums still require validation through sustained product execution, demand conversion, and margin improvement.

Risks

• Intensifying competition in the global foundation-model industry, with further narrowing model performance gaps.
• High R&D investment could delay earnings visibility and increase cash burn and self-funding pressure.
• Progress in model capabilities, commercialization conversion, and non-inference revenue may fall short of expectations.
• Escalating China-U.S. technology competition creates geopolitical and supply-chain uncertainty.
• Positive scenarios include better-than-expected model intelligence, a faster-clearing profitability path, additional non-inference revenue-sharing income, and better-than-expected commercialization capabilities.

What to watch

• Launches and performance of GLM-5.5 and 2 trillion to 5 trillion parameter models from other Chinese model vendors.
• Developer adoption following the release of GLM-5.3 model weights, coding benchmark performance, and per-task costs.
• ZCode active users, off-peak task utilization, and post-training benefits generated from product-data feedback.
• The impact of GLM Coding Plan subscriptions, API pricing, and industry price adjustments on revenue and gross margin.
• Whether ASIC and inference-chip capacity ramp-up drives conversion of open-source demand into ARR.
• Market share, long-term margins, cash burn, and the transition from open source to open-weight models with commercial terms attached.