GLM-5.3 Capability Upgrade Reinforces the Post-Training Scaling Thesis, but Valuation and Competitive Pressure Keep Rating Neutral

Overview

This report focuses on Z.AI Co.'s technology and commercialization progress following the launch of GLM-5.3. Goldman Sachs believes this release validates the importance of post-training scaling for improving the capabilities of Chinese large models and strengthens the company's competitive positioning in coding models, agent task execution, and the open-source ecosystem. Despite the long-term fundamental opportunity, the report maintains a Neutral rating due to valuation, competition, and earnings uncertainty.

Core views

GLM-5.3's performance improvement is primarily driven by post-training scaling rather than an increase in parameter count, reflecting the benefits of longer task environments, longer training cycles, and upgraded reinforcement learning frameworks.The company leads in open-source coding models, with model weights planned for release within two weeks; its MIT license contains no commercial restrictions or revenue-sharing terms.ZCode is both a distribution and monetization tool and an important data entry point for obtaining long-horizon, real-world coding trajectories, supporting subsequent model post-training.Subscription and API price changes indicate that improved model capabilities are strengthening pricing power and moving the industry from aggressive subsidies toward more rational pricing norms.Goldman Sachs believes Z.AI's valuation premium has some fundamental justification, but narrowing capability gaps with peers and limited near-term profit visibility keep the current risk-reward profile balanced.

Analysis framework

The report evaluates technological progress through model benchmark testing and product-function observations, compares the parameter scale, pricing, and valuations of major Chinese large-model vendors, and derives the target price using a DCF framework while assessing risk-reward under bull, base, and bear scenarios.

Methodology notes

• Valuation methodsDiscounted Cash Flow Valuation

  DCF

  Valued using a 12% weighted average cost of capital and a 2% perpetual growth rate, incorporating assumptions for market share and long-term adjusted EBIT margin.

• Model TrainingPost-Training Scaling

  Enhancing model capabilities through longer task environments, training cycles, and reinforcement learning upgrades

  The report attributes GLM-5.3's capability improvement to post-training scaling rather than expanding its 744 billion parameter count.

• Competitive AnalysisPerformance-Price Comparison

  Comparing model capability, cost per task, API pricing, and ARR valuation multiples

  Used to assess Z.AI's technological advantages, commercialization potential, and valuation reasonableness relative to DeepSeek and MiniMax.

Asset mapping & comparison

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

• Z.AI Co. (2513.HK)

  Covered Company

  Strengths

  GLM-5.3 has improved notably in coding, long-horizon tasks, and cybersecurity; ZCode's user scale and task data support a post-training data flywheel; the model offers competitive performance-price value.

  Weaknesses

  Near-term profit visibility is limited, while high R&D spending and cash-burn pressure remain significant; model-performance gaps may narrow.

  Comparison

  The company has a market capitalization of approximately US$75bn, equating to around 30x Goldman Sachs' forecast of US$2.5bn in end-2026 ARR; DeepSeek was recently valued at approximately US$50bn, while MiniMax has a market capitalization of approximately US$14bn and trades at roughly 14x ARR.

  Risks

  Global foundation-model competition, weaker-than-expected commercialization, inference-capacity constraints, self-funding capability, and intensifying China-U.S. technology competition.

• DeepSeek

  Key Competitor

  Strengths

  Has large-parameter-scale models and industry influence; its recent increase in V4 API pricing reflects an improving pricing environment.

  Weaknesses

  The report states that GLM-5.3 outperforms DeepSeek V4 Pro on multiple benchmarks.

  Comparison

  DeepSeek V4 Pro 0813 peak-period blended pricing is approximately US$0.69 per million tokens, close to GLM-5.2's US$0.88.

  Risks

  Industry competition and pricing changes may affect Z.AI's relative advantages and valuation.

Key data

• RatingNeutralMaintained unchanged.
• 12-Month Target PriceHK$1,610Based on DCF valuation.
• Current PriceHK$1,270Price shown in the report table.
• Implied Upside26.8%Calculated from the current price and 12-month target price.
• GLM-5.3 Parameter Scale744 billion parametersUnchanged from GLM-5.2; performance improvement primarily comes from post-training scaling.
• ZCode Active UsersMore than 1 millionThe report views this as an important foundation for distribution, monetization, and the training-data flywheel.
• Coding Plan Lite PriceRMB 118/monthThe company states that, based on weekly token allowances, this represents an approximately 50% discount to equivalent API costs.
• Base-Case Market Share21% in 2030DCF valuation assumption.
• Long-Term Adjusted EBIT Margin26% in 2035DCF valuation assumption.
• Bull/Bear Case Implied ValuationHK$3,266 / HK$835Based on the different key-assumption scenarios presented in the report.

Impact & implications

GLM-5.3 indicates that the key variables in China's large-model competition are shifting from simply scaling parameter counts toward post-training efficiency, task environments, inference costs, and real-world usage data. If ZCode continues to expand its user base and task coverage, the company could build a stronger data flywheel and unlock model demand and ARR growth as future inference capacity improves. However, narrowing model-performance gaps and high R&D investment mean that technological leadership still needs to translate into verifiable revenue and profits.

Risks

• Intensifying competition in the global foundation-model industry, with narrowing model-capability gaps among peers.
• High R&D expenses resulting in insufficient near-term earnings visibility.
• Uncertainty regarding cash burn and self-financing capacity.
• Escalating China-U.S. technology competition creating geopolitical and supply-chain risks.
• Model commercialization, non-inference revenue-sharing income, or profit improvement may fall short of expectations.
• Model intelligence, commercialization capabilities, and the profitability path could also exceed expectations, representing upside risk relative to this Neutral view.

What to watch

• The launch and actual performance of GLM-5.5 and 2-trillion- to 5-trillion-parameter models from other Chinese vendors.
• Developer adoption, coding benchmark performance, and inference costs following the release of GLM-5.3 open-source weights.
• ZCode user growth and usage of cross-device task management and multi-agent collaboration features.
• The impact of off-peak task functionality on inference-capacity utilization, user costs, and task-data accumulation.
• Whether model subscription and API pricing, as well as industry price competition, continue to move toward rationalization.
• Progress in ARR growth, inference-chip capacity ramp-up, R&D spending, cash burn, and profit improvement.