Chinese-Made MoE LLM Put to the Test: 3-Trillion-Parameter Model Nears Top International Performance at One-Fifth the Price

Introduction: The LLM Race Enters a Price-Performance Era

In late October 2025, a major Chinese AI company launched its new-generation flagship model, officially described as “the most capable model the company has ever built.” Unlike the usual industry pattern of publishing reports before releasing a product, this model first appeared anonymously on third-party evaluation platforms, where developers tested it without knowing its identity. The company only revealed the name on launch day. That reversed process reflects the balance Chinese LLM vendors are now striking between technical confidence and market competition.

In terms of scale, the model uses a Mixture of Experts (MoE) architecture with roughly 3 trillion total parameters, reported by industry media as 2.8 trillion. Each inference activates only about 104 billion parameters. The context window extends to 1 million tokens, enough to process a medium-sized code repository or a long technical document in a single pass. A high-speed variant was released alongside the standard version, offering the same intelligence level with output speed improved by 2 to 3 times.

01 Architecture Breakdown: How MoE Delivers Better Performance at Lower Cost

1.1 The Economics of Sparse Activation

A traditional dense model has to engage all parameters during inference, making compute cost scale almost linearly with model size. The MoE architecture takes a different approach: it splits the model into multiple expert sub-networks and activates only a small subset for each inference. In the case of this newly released model, only about 3.5% of the 3 trillion total parameters are activated per query. That means a single A100 or H100 GPU can handle inference, instead of requiring clusters with hundreds of cards.

This mechanism directly shapes the pricing. The model’s API price is reported to be only one-fifth that of a major international closed-source flagship. For enterprise customers making millions of API calls per day, that could reduce annual inference costs from tens of millions to a few million yuan, giving it a clear price-performance advantage.

1.2 Engineering Trade-offs in Long Context and Multimodality

The 1-million-token context window is not simply the result of adding more GPU memory and more parameters. It requires systematic optimization across attention mechanisms, KV cache compression, and long-text positional encoding. In testing, the model can continuously read and understand roughly 750,000 characters of a codebase or technical manual, then accurately cite details from that material in later responses. This has direct value for code review, document Q&A, and cross-file refactoring.

The model also natively supports visual input, meaning it can understand screenshots, flowcharts, UI design mockups, and similar content. It scored 92.7% on BenchCAD, a 3D reconstruction benchmark. That still trails leading closed-source models at 95.9%, but it is already ahead of most models in the open-source community.

02 Benchmarks: Coding Agent Emerges as the Strongest Area

2.1 Key Benchmark Comparison

Official evaluation results compare the model with two major international closed-source flagships across several authoritative benchmarks:

Benchmark	This Model	GPT-6 Astra	Claude Opus 5.5
Terminal-Bench 4.0	65.7%	58.2%	66.4%
SWE-Marathon v1.1	51.7%	42.5%	46.3%
Agents’ Last Exam	43.2	32.8	42.6
BenchCAD (3D reconstruction)	92.7%	95.9%	96.2%

2.2 Coding Agent Capabilities

Terminal-Bench 4.0 measures how well a model performs multi-step operations in a real terminal environment, including running commands, reading output, modifying files, and handling errors. The model scored 65.7%, well ahead of one international flagship at 58.2% and only 0.7 percentage points behind the highest score of 66.4%. This is a meaningful result because terminal evaluation reflects a model’s ability to close the loop from intention to action, which is more relevant than pure code generation for judging whether an AI coding assistant can work in practice.

SWE-Marathon v1.1 focuses on long-running software engineering tasks such as cross-file feature development and multi-branch merging. The model scored 51.7%, beating both strong reference models. That suggests it can maintain context coherence and decision quality during task chains lasting more than 30 minutes, rather than scoring well on short tasks and degrading over longer ones.

2.3 Third-Party Evaluation

Independent evaluation organization Artificial Analysis gave the model an intelligence index score of 55, ranking fourth globally and 11 points higher than the previous generation. The organization also predicted that after the model weights are released, the average score gap with leading closed-source models could shrink from 12 points to 3 points, nearly erasing the traditional divide between open-weight and closed-source models.

03 Weaknesses and Limitations: Gaps That Cannot Be Ignored

3.1 Three Clear Weak Spots

Despite the strong results, the model still has three notable limitations:

• Scientific computing terminal tasks: Performance falls below average in scenarios involving complex formula derivation and specialized notation. This likely stems from insufficient coverage of dense scientific literature in the training data.
• Limited depth in multimodal understanding: Although the model supports visual input, instruction tuning still has room to improve in fine-grained visual reasoning, such as detecting subtle differences in charts and diagrams.
• Stability under high concurrency: The high-speed variant occasionally shows response latency fluctuations at peak load, which requires further optimization at the engineering and scheduling layer.

3.2 The Real Gap with Leading Closed-Source Models

Overall, the model is already close to, and in some cases ahead of, closed-source flagships in coding and agent tasks. However, it still lags by roughly 2 to 4 percentage points in knowledge breadth, complex logical chain reasoning, and few-shot learning consistency. That means leading closed-source models still hold an advantage for tasks that depend heavily on world knowledge or strict logical reasoning, such as academic paper review or legal text analysis.

04 Three Recommendations for Developers and Enterprise Users

4.1 Clarify Requirements Before Migrating

• Well-suited scenarios: code generation and review, long-document understanding, API orchestration and agent development, and data analysis report generation.
• Use with caution: high-precision scientific computing, decision support in heavily regulated compliance contexts, and applications requiring extremely long multi-turn conversations with strict consistency verification.

4.2 Watch for Self-Hosting Opportunities After Weight Release

According to reliable reports, the model weights will be released later. Enterprises will then be able to deploy the model on private clouds or their own infrastructure and combine it with RAG (retrieval-augmented generation) and fine-tuning to build dedicated AI systems. This would further reduce marginal costs and help address data compliance risks.

4.3 Use Test-Driven Model Selection, Not Benchmark Worship

A more reliable approach is to validate against your own data:

1. Select 100 to 300 real business samples from your organization.
2. Call multiple candidate models in parallel for blind A/B testing.
3. Have business experts score the results on accuracy, robustness, and response speed.
4. Combine those results with API pricing and projected call volume to perform a total cost of ownership (TCO) analysis.

Conclusion: A Watershed Moment

The release of this 3-trillion-parameter MoE model makes the case with concrete cost and performance data that LLM competition is shifting from a parameter race to a pragmatism race. When the gap between open-weight models and closed-source flagships narrows to about 3 percentage points while the price differs by a factor of five, enterprise selection logic will be fundamentally reshaped.

For developers, this means lower barriers to experimentation, more flexible technology choices, and stronger autonomy and control. Over the next few years, the democratization of AI infrastructure will accelerate, and each leap in price-performance will create conditions for entirely new application categories.