GLM-5.3 Strengthens Zhipu's Competitiveness in Coding, Agents, and Cybersecurity

Overview

The report focuses on the technological and investment implications following Zhipu's release of GLM-5.3. Bernstein believes the model delivers meaningful improvements in long-horizon coding, agent execution, automation, and cybersecurity, while enhancing competitiveness at lower Token pricing.

Core views

GLM-5.3 could return Zhipu to the frontier with a balance of performance and cost. Relative to Qwen3.8 Max, the report views its broad agent capabilities as modestly improved, while its advantages in the coding-to-agent pipeline and cybersecurity are more pronounced; it still trails the most advanced U.S. models on the most challenging software-engineering tasks. Stronger cybersecurity capabilities could expand the total addressable market and support the subsequent growth narrative.

Analysis framework

The report assesses model performance, Token efficiency, competitive positioning, and investor sentiment using company-disclosed benchmarks, Bernstein proprietary tests, a 24-hour post-release social media sample, and developer feedback; valuation combines 2030 forward P/E with discounted cash flow analysis.

Methodology notes

• Model EvaluationBenchmark Comparison

  Compares model capabilities through test scores across coding, agents, cybersecurity, and other areas.

  The report cross-validates GLM-5.3's performance using company-disclosed results and proprietary tests, and compares it with Qwen, U.S. frontier models, and other competing models.

• Market SentimentSentiment Cycle Heuristic

  Uses market feedback and share-price movements after the model release to assess tactical sentiment risks.

  The report believes market expectations for AI lab stocks change rapidly, and the sentiment cycle will continue to affect near-term performance after the recent share-price increase.

• Valuation methodsP/E and Discounted Cash Flow

  Values the company by discounting forward earnings multiples and combining them with discounted cash flow analysis.

  Bernstein values Zhipu using 25x 2030 forward P/E, a 14% annual discount rate, and DCF.

Asset mapping & comparison

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

• 2513.HK

  Zhipu listed equity

  Strengths

  GLM-5.3's improved agent, coding, and cybersecurity capabilities, together with unchanged pricing and improved Token efficiency, benefit product competitiveness and commercialization expansion.

  Weaknesses

  It still trails the most advanced U.S. models on the most challenging long-horizon software-engineering tasks and remains loss-making during the forecast period.

  Comparison

  The report considers it ahead of Qwen3.8 Max in the coding-to-agent pipeline and cybersecurity; it remains behind the most advanced U.S. models, but leads Opus 4.8 on most agent and cybersecurity benchmarks.

  Risks

  The pace of enterprise AI adoption, peer-model competition, weakening AI sentiment, and debate over long-term valuation could all intensify share-price volatility.

Key data

• Investment RatingOutperformBernstein rating, based on expected relative performance versus the benchmark over 12 months.
• Target PriceHK$1,350Represents approximately 6% potential upside from the HK$1,270 closing price on 2026-08-14.
• GLM-5.3 Parameter ScaleApproximately 750 billion total parameters and 40 billion activated parametersUses the same pre-training backbone as GLM-5.2.
• CyberGym Score84.5The report lists Mythos 5 at 83.8 and GPT-5.6 Sol at 83.6.
• Vulnerabilities IdentifiedMore than 2,400Pre-release testing covered 269 projects, with approximately 1,100 medium- to high-severity issues.
• Social Media Feedback85%Among 1,000 sampled Twitter/X posts in the 24 hours after release, 85% of the 419 posts expressing directional views were positive.
• Expected Revenue for 2026 and 2027RMB 5,631 million and RMB 18,663 millionForecast values in the report's table.

Impact & implications

Technological progress supports a stronger combination of performance, efficiency, and price for Zhipu in the mid-market, and creates an opportunity to develop cybersecurity into a third commercialization pillar. If a subsequent, larger-scale pre-trained model launches as expected, it could further narrow the gap in the most difficult long-horizon software-engineering tasks; however, market disagreement over valuation based on long-term growth remains substantial.

Risks

• Chinese enterprise customers may adopt AI more slowly than expected.
• Competitors' model iterations may erode technological and pricing advantages.
• Investor views on Chinese AI labs may shift rapidly with the latest model releases and market sentiment.
• Valuation based on 2030 growth expectations may remain contentious.
• The model's lack of vision capabilities may limit certain application scenarios.

What to watch

• Follow-up testing results for GLM-5.3 from third-party platforms.
• Cybersecurity partner networks, vulnerability-testing data, and commercialization progress.
• The launch timing of Zhipu's next-generation, larger-scale pre-trained model, for which management guidance is September to October.
• Token price trends and competitors' pricing changes.
• Conversion of enterprise AI demand, revenue growth, and improvement in losses.
• AI industry sentiment and share-price volatility in 2513.HK.