A Trillion-Parameter Chinese LLM Put to the Test: Near-Frontier Performance at One-Fifth the Cost

A New Benchmark for Chinese LLMs: Breaking Through on Both Parameters and Performance

As large language model competition intensifies, vendor announcements have shifted from parameter counts toward benchmark results and cost-effectiveness. A major Chinese AI company recently launched its latest flagship model, describing it as the most capable model the company has ever produced. The model joins the trillion-parameter club while delivering benchmark results that approach or, in some cases, surpass leading closed-source international models.

This article examines the model from four angles: technical architecture, measured performance, pricing strategy, and ecosystem impact.

1. Parameter Scale and Technical Foundation: The Reality Behind the “3 Trillion” Figure

The most eye-catching label attached to this model is “3 trillion parameters.” However, this is not a dense model in the traditional sense. According to the company and multiple industry reports, the model uses a Mixture of Experts (MoE) architecture, with a total parameter count of roughly 2.8 to 3 trillion. During each inference pass, only about 104 billion parameters are actually activated.

This design is key to understanding the model’s performance and cost profile. The core idea behind MoE is specialization: the model is split into multiple expert sub-networks, and each input is routed only to the most relevant subset of experts. This delivers two direct benefits:

1. Significantly lower inference costs: Compared with a dense model of similar scale, where every inference pass engages all parameters, the MoE model activates only about 3% to 5% of its total parameters. This substantially reduces compute consumption per token and underpins the vendor’s aggressive pricing.
2. Decoupling model capacity from inference cost: A larger total parameter count allows the model to store broader world knowledge and reasoning patterns, while sparse activation preserves runtime efficiency. This breaks the traditional assumption that larger models are always slower to run.

Beyond parameter scale, the model supports a context window of up to 1 million tokens. That capacity is enough to process a mid-sized code repository, a full-length novel, or dozens of pages of technical documentation in a single pass, laying the foundation for complex agent tasks and deep document analysis. The model also natively supports vision understanding and can process image inputs directly.

Notably, before the official launch, the model was tested anonymously on third-party evaluation platforms under a placeholder name. The vendor only claimed the results on release day. This “evaluate first, announce later” approach contrasts with the traditional pattern of publishing benchmark reports before a product is widely available, and it signals greater attention to real-world evaluation feedback.

2. Benchmark Analysis: Strong Terminal Agent Performance, with Gaps in Specialized Domains

Raw parameter counts do not define model quality. The more meaningful question is how the model performs on real tasks. According to third-party benchmark data cited by the vendor, the model is highly competitive in several areas, but it is not uniformly ahead.

1. Coding and Agent Tasks: A Leading Position

On Terminal-Bench 4.0, which evaluates coding agent capabilities in real terminal environments, the model achieved 65.7%, well ahead of one major international model at 58.2% and only 0.7 percentage points behind another closed-source flagship at 66.4%. This suggests the model performs at a top international level on complex agent tasks such as instruction understanding, tool calling, code execution, and error handling.

On SWE-Marathon v1.1, which assesses long-horizon software engineering tasks, the model scored 51.7%, surpassing two major international competitors, which scored 46.3% and 42.5% respectively. This points to strong capabilities in long-chain reasoning, state maintenance, and multi-step planning.

The model also ranked first on Agents’ Last Exam, a comprehensive agent evaluation. This further supports its generalization across agent tasks rather than being limited to narrow coding scenarios.

2. Specialized Domains: Some Gaps Remain

Despite strong results on coding and agent benchmarks, the model scored 92.7% on BenchCAD, a 3D reconstruction benchmark, trailing two competitors at 95.9% and 96.2%. International frontier models also retain advantages in certain scientific computing terminal tasks and some deep knowledge reasoning scenarios. This indicates that the model’s current strengths are concentrated in engineering and logical reasoning, while areas requiring specialized domain knowledge or fine-grained visual-spatial understanding still have room to improve.

3. Independent Assessment: A Promising Step for Open-Weight Models

Independent research organization Artificial Analysis gave the model an intelligence index score of 55, ranking it fourth globally. That is an 11-point jump from the previous generation’s score of 44. The organization specifically noted that if the model’s weights were opened, it would immediately become the highest-scoring open-weight model available, narrowing the average intelligence gap between open-weight and closed-source models from 12 points to 3 points. For the open-source community and small-to-medium enterprises, this means access to capabilities that previously existed only inside closed-source frontier labs.

3. Pricing Strategy: Cost Disruption and Industry Restructuring

One of the most disruptive aspects of the launch is that the model is priced at roughly one-fifth of a leading international flagship model. Behind this pricing is the cost advantage of the MoE architecture, as well as a clear intent to capture market share and lower the barrier for developers.

For enterprise users, API cost is a core factor in determining whether a large language model can be deployed at scale. A fivefold price reduction means:

• Lower experimentation costs: Teams can run extensive experiments, fine-tuning, and prompt engineering without losing control of their budgets.
• Broader application scenarios: In cost-sensitive use cases such as log analysis, content moderation, and batch information extraction, affordable high-performance models make AI-based approaches viable replacements for traditional rule-based systems.
• New business models: Lower base costs give SaaS providers and AI-native application developers more room to build sustainable margins.

This is not an isolated case. Several Chinese model vendors are now competing on cost-effectiveness against international giants. A broader trend is emerging: competition in foundation models is shifting from a pure capability race toward a combined capability-per-unit-cost race. For downstream application developers, this is arguably the best environment to date.

4. From Technology to Ecosystem: Opportunities and Challenges for Developers

For developers and architects, adopting a model of this class requires planning across both the existing tech stack and project architecture.

1. Architecture adaptation: Pay attention to API compatibility and tooling
   Even with top-tier model capabilities, introducing a new model usually requires adapting existing prompts and tool-calling protocols. Developers should consider adding a model-independent abstraction layer early in a project so that different vendors can be plugged in through a unified interface. This allows flexible switching based on performance and cost.

2. Feature selection: Prioritize long context and agent capabilities
   The 1-million-token context window enables use cases that were previously impractical. For example, the model can read an entire code repository to support cross-file bug localization and refactoring suggestions, or process a full industry white paper in one pass and generate a deep analysis report. Its strong agent capabilities also make it possible to build automated workflows such as generating test cases directly from requirements.

3. Cost optimization: Use a hybrid scheduling strategy
   Not every request needs to be routed to the most powerful model. A sensible architecture uses the flagship model for the most complex, highest-value reasoning tasks while directing simpler jobs such as text classification or entity extraction to lighter models. This model-routing approach can preserve quality while further reducing total cost.

5. Conclusion: What This Means for the Next Phase of Competition

The release of this “3-trillion-parameter” model is not just another benchmark announcement. It marks a maturing Chinese LLM industry and sends three clear signals:

• The technology gap is narrowing: In key engineering and agent capabilities, top Chinese models can now compete head-to-head with the international frontier.
• Cost competition is becoming a primary battleground: Compute efficiency is now a core competitive advantage. Future models will need to be affordable enough for broad adoption.
• Open weights could be the ultimate catalyst: If strong model weights are opened, the global developer community could unlock a new wave of AI application growth.

For developers, the practical takeaway is to follow frontier model progress closely, keep the tech stack flexible and portable, and evaluate each model’s cost-effectiveness in the context of specific business scenarios. The window for turning large language models from experimental tools into production infrastructure is opening quickly.