Open-weight commercialization has entered a 'stronger-wins-more' phase

Overview

This report discusses the commercialization path of Chinese large language model providers under open-weight models. Its core view is that open weights should not be understood simply as leakage of first-party API revenue, but as a change in the revenue mix: while vendors may lose some direct model-access income, they can gain new paid conversion through broader distribution, official APIs, cloud partnerships, enterprise deployment, revenue sharing, SaaS markets, and workflow products.

Core views

The report presents a "winner-takes-more" framework: the stronger the model capability, the easier it is to convert wider distribution from open weights into commercial upside. Zhipu AI benefits because GLM-5.2 has a stronger competitive positioning, with open weights increasing adoption and flowing back to official channels and premium service tiers. Although MiniMax is pursuing a similar strategy, evidence of M3 differentiation and model-led pricing power is weaker, making it more exposed to substitution and price competition.

Analysis framework

The analysis is structured across four dimensions: first, comparing quality and service differences between open-weight and official API paths; second, using DeepSeek V4 Pro and MiniMax M3 as case studies to show pricing, caching, and quality differences across providers for the same model; third, analyzing the trade-off CSPs and inference platforms face between self-deployment, which preserves margin, and official partnerships, which help protect quality; fourth, discussing how coding agents, enterprise workflows, and first-party harnesses can convert open-model adoption into stickier paid relationships.

Methodology notes

• Commercialization frameworkOpen-weight winner-takes-more framework

  Open weights expand distribution, but monetization concentrates in model providers with leading capability, fast iteration, reliable endpoints, and workflow embedding.

  Strong models can increase adoption via open weights and capture value through official APIs, cloud partnerships, enterprise deployment, and workflow products. Weaker models are more easily deployed by third parties, arbitraged, and substituted.

• Channel analysisMargin and quality trade-off

  CSPs and inference platforms choose between self-deployment and official partnerships.

  Self-deployment can retain more gross margin but requires taking responsibility for model updates, quantization, caching, reliability, and customer support; official partnerships reduce platform share but improve freshness, technical support, and enterprise trust.

• Product strategyFirst-party workflow harness

  Model providers can own user relationships through coding agents or workflow products.

  If users consume models within provider-owned products, model companies can capture direct feedback, data, retention, and higher API stickiness, reducing the endpoint substitutability caused by open weights.

Asset mapping & comparison

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

• Zhipu AI

  Core beneficiary

  Strengths

  GLM-5.2 is viewed as having global competitive positioning; open weights can expand adoption; official pathways and high-service versions such as GLM-Turbo can meet quality-sensitive demand; target price is raised and Overweight is maintained.

  Weaknesses

  Revenue pathways still depend on sustained model leadership, and the market has already largely priced in the US$1bn year-end ARR target.

  Comparison

  Compared with MiniMax, the report sees stronger evidence that Zhipu has better model competitiveness and greater monetization optionality under open weights.

  Risks

  Whether GLM-5.2 represents a company-wide capability inflection, competitive performance of Kimi K3 and DeepSeek V4.1, and whether GLM-5.5/6 can sustain the advantage.

• MiniMax Group Inc - H

  Under pressure

  Strengths

  M3's official route may outperform some third-party routes in cache hit rate, realized input cost, usage concentration, and speed.

  Weaknesses

  M3 has not yet shown sufficient model differentiation or pricing power; open weights may accelerate arbitrage, routing, and substitution.

  Comparison

  Compared with Zhipu, the report judges MiniMax's model-led conversion-to-revenue evidence as weaker, so the target price is reduced and Neutral is maintained.

  Risks

  If model capability, official endpoint quality, or first-party workflow products improve, paid conversion could be better than expected; otherwise competitive and pricing pressure may continue to rise.

• DeepSeek

  Case reference

  Strengths

  The DeepSeek V4 Pro official API forms a pricing moat through low list pricing and caching economics.

  Weaknesses

  Open weights still allow third-party deployment, which diverts some traffic.

  Comparison

  As a price-led official-route case, it shows that open weights do not necessarily commoditize first-party API revenue.

  Risks

  If third-party routes catch up in quality and cost, official API advantages may narrow.

Key data

• Zhipu AI target priceHK$2,000The Dec-26 target price was raised from HK$1,800 to HK$2,000, with Overweight maintained.
• MiniMax target priceHK$300The Dec-26 target price was cut to HK$300, with Neutral maintained.
• DeepSeek V4 Pro official-route costaround US$24-41Under a simplified monthly workload of 100M effective input tokens and 20M output tokens, the official route is cheaper than some third-party routes at about US$85-196.
• DeepSeek V4 Pro official-route price advantageabout 6-12x cheaperThe advantage is driven by lower list pricing and stronger caching economics.
• Zhipu AI FY26E revenueRmb 4,715mnForecast disclosed in the JPMorgan table.
• MiniMax FY26E revenueUS$383mnForecast disclosed in the JPMorgan table.

Impact & implications

The report's investment implication is that open weights will accelerate outbound expansion of Chinese large models and cloud distribution, but commercialization gains will be more concentrated. Investors should distinguish whether models genuinely have SOTA or near-SOTA capability, whether official endpoints consistently outperform third-party deployment, and whether providers can own customer relationships through workflow products. This framework supports a more constructive valuation for Zhipu AI while placing greater valuation pressure on MiniMax when model differentiation is not clearly demonstrated.

Risks

• Leading model capability does not persist, and the adoption gains from open weights fail to translate into paid revenue.
• Third-party clouds, inference platforms, or API aggregators provide acceptable quality at lower cost, causing first-party API traffic to be rerouted.
• If official endpoints do not maintain advantages in caching, long context, tool calling, latency, and stability, premium pricing power may weaken.
• Competition at coding-agent and workflow entry points in China is intense; products such as Tencent WorkBuddy, ByteDance Trae, and Alibaba Qoder may limit the migration room for model provider first-party harnesses.
• Mature tasks like simple chat, summarization, translation, and basic Q&A continue to see falling prices, compressing generic API revenue.

What to watch

• Whether GLM-5.2 reflects a company-wide capability step-up for Zhipu, and whether subsequent GLM-5.5/6 can widen the gap.
• Capability and pricing comparisons among Kimi K3, DeepSeek V4.1, and GLM-5.2.
• How channels such as AWS Bedrock, AWS Marketplace, Azure Foundry, Fireworks AI, Cloudflare, and OpenRouter distribute China open-weight models.
• Whether official APIs and third-party endpoints diverge in cache hit rate, effective input cost, latency, long-context handling, and tool-calling quality.
• Whether first-party workflow products like Z Code, MiniMax Code, and Kimi Code can create user retention and paid conversion.