The wave of open-sourcing model weights is reshaping the competitive landscape of China's large language model (LLM) market. In a recent report, JPMorgan notes that the open-source strategy is evolving into a "winner-takes-most" business model. The stronger the model's capabilities, the more effectively it can convert open distribution into paid monetization. Conversely, models lacking sufficient differentiation face heightened risks from price competition and traffic diversion.

Based on this analysis, JPMorgan has issued divergent ratings for two Chinese AI companies. The report raises the target price for KnowledgeAtlas (ASX: 02513) to HK$2,000 from HK$1,800 for December 2026, maintaining an "Overweight" rating. This adjustment is driven by the belief that the GLM-5.2 model strengthens the thesis that open-weight commercialization can create significant option value for leading model providers. Simultaneously, the target price for MiniMax is lowered to HK$300 from HK$400, with a "Neutral" rating retained, as its M3 model has yet to demonstrate sufficient model-driven pricing power.

Open Weights: A Structural Shift in Monetization Logic

JPMorgan suggests that while the market commonly views the release of open weights as a potential drain on monetization, this perspective is correct but incomplete. The more critical question is whether LLM providers can transform weakened access control into broader distribution and paid conversion.

The report highlights that open-weight releases do not make all API endpoints identical. Official APIs retain systemic advantages in areas such as model timeliness, caching strategies, throughput, latency, feature support, and service reliability.

An open-weight release is typically a static checkpoint, whereas an official API is a continuously evolving product. After release, providers can optimize their models based on real-time traffic for instruction-following, tool use, long-context stability, and inference efficiency. These optimizations are not always fully synchronized with the open-weight package. Consequently, the actual user experience between the official endpoint and third-party deployments under the same model name may diverge over time.

KnowledgeAtlas: GLM-5.2 Strengthens Leading Position, Target Raised to HK$2,000

JPMorgan maintains its "Overweight" rating on KnowledgeAtlas and raises the target price, with the core logic being that GLM-5.2 enhances the potential for open-weight monetization.

The report views KnowledgeAtlas's open-weight strategy as more measured: using relaxed access to expand usage scale, while positioning its official path and higher-service versions (like the GLM-Turbo series) to meet quality-sensitive demands. GLM-5.2, released under an MIT license, continues to expand its distribution footprint across global cloud and inference providers.

JPMorgan notes that, based on current valuations, the market has largely priced in KnowledgeAtlas's guidance for $1 billion in annual recurring revenue by year-end. The remaining upside potential hinges on whether its robust open-weight models can achieve scale through external infrastructure and distribution channels, rather than relying solely on its own GPU compute stack. This represents option value rather than guaranteed near-term revenue and can only scale if KnowledgeAtlas maintains its model leadership. Key metrics to watch include the comparative performance of GLM-5.5/6 against rivals like Kimi K3 and DeepSeek V4.1.

MiniMax: Lack of Pricing Power Amplifies Downside Risks from Open Weights

JPMorgan maintains a "Neutral" rating on MiniMax, lowering its target price, citing that open-weight commercialization is becoming a "winner-takes-most" framework, and the M3 model has not yet shown sufficient evidence of model-driven pricing power.

The report views the M3 model's permanent 50% discount as a significant signal that it has not achieved a capability premium against leading domestic models. While the discount may support short-term usage volume, it diminishes market confidence in model-driven monetization. In an environment of expanded access via open weights, models with insufficient differentiation face faster price comparison and easier traffic diversion, turning broader distribution into a risk factor rather than a source of upside.

JPMorgan also acknowledges MiniMax's relative strengths, such as its relevance in multimodal AI, overseas usage, and agent workflows. The M3 model has improved its product narrative with features like a 1 million token context window, native multimodality, and MiniMax Code. Adoption data from platforms like OpenRouter indicates significant developer usage. However, monetizing workflows requires stronger model pull—products for coding or agents need to offer substantial improvements in task completion to change user habits, not just provide another access path to a widely available model.

Key Risks and Capital Needs

JPMorgan also cautions that both companies are in capital-intensive stages, with financing needs posing a potential risk. Analysts anticipate that both companies may require two additional funding rounds in 2026 and 2027. While current cash balances are sufficient to support operations under base-case scenarios, accelerated model iteration, larger-scale overseas deployment, or higher-than-expected inference costs could prompt both firms to seek additional external capital.