A Flagship Chinese MoE Model Put to the Test: 3 Trillion Parameters, Near-Top-Tier Closed-Source Performance, and Inference Costs Cut to One-Fifth

From Parameter Scale to Engineering Reality: A Full Breakdown of a Flagship Chinese MoE Model

A Chinese AI vendor recently released its most capable flagship large language model to date. With roughly 3 trillion total parameters, a 1 million token context window, and aggressive pricing, it quickly drew attention from the developer community. In several third-party benchmarks, the model approaches or even surpasses leading overseas closed-source models in key areas, while its price is only about one-fifth that of comparable flagship models. This article provides a systematic breakdown of the model across four dimensions: technical architecture, benchmark analysis, cost logic, and scenario fit.

01 Architecture Analysis: The 3 Trillion Parameter Strategy and MoE Engineering

The first concept to clarify is that 3 trillion does not refer to dense activated parameters in the traditional sense. Instead, it is the total parameter count of a Mixture-of-Experts (MoE) architecture. Public information indicates that the model has between 2.8 trillion and 3 trillion total parameters, while each inference pass activates roughly 104 billion parameters. This design reflects an important trend in large model engineering: using sparse activation to improve inference efficiency.

The core advantage of MoE is that it splits a large neural network into multiple expert sub-networks. During each inference step, only a small subset of experts is dynamically activated based on the input tokens. As a result, although the model theoretically holds the knowledge capacity of 3 trillion parameters, its actual computational cost is comparable to a dense model in the hundreds-of-billions range. This “large but not slow” characteristic directly enables the vendor to offer the model at a very low price: per-token inference cost is substantially reduced, while the model retains a capacity ceiling close to much larger dense models.

The expansion of the context window to 1 million tokens is another major hardware-level improvement. This allows the model to process medium-sized code repositories, hundreds of pages of technical documentation, or long multi-turn conversation histories in a single pass, without requiring developers to build complex chunking and retrieval logic. Combined with native vision understanding, the model has natural advantages in mixed text-image document parsing, UI automation testing, and converting code screenshots into text.

The release process is also noteworthy. Before the model received its official name, the team made it available on third-party blind evaluation platforms under an anonymous codename. Only after collecting real-world feedback did the vendor publicly claim the results. This “enter the arena first, reveal the name later” approach contrasts sharply with traditional product launch patterns and places greater emphasis on real-environment testing than on internal evaluation data alone.

02 Benchmark Deep Dive: Strong in Coding and Agent Tasks, Weaker in Geometric Reasoning

According to the official benchmark results, the model shows a clear performance profile across different dimensions. The table below compares key metrics, with higher scores indicating better performance:

Benchmark	This Model	Overseas Flagship A	Overseas Flagship B
Terminal-Bench 4.0 (terminal/shell operations)	65.7%	58.2%	66.4%
SWE-Marathon v1.1 (multi-turn code repair)	51.7%	42.5%	46.3%
Agents’ Last Exam (general agent evaluation)	43.2	32.8	42.6
BenchCAD (3D scene reconstruction)	92.7%	95.9%	96.2%

Key takeaway: the model excels in code generation and agent tasks, approaching top-tier closed-source models.

• Terminal-Bench 4.0: This benchmark simulates a real terminal environment and requires the model to perform command-line operations, script writing, and environment configuration. The model scored 65.7%, only 0.7 percentage points behind Overseas Flagship B (66.4%) and well ahead of Overseas Flagship A (58.2%). This indicates strong practical value in system operations automation, CI/CD pipeline assistance, and related scenarios.
• SWE-Marathon v1.1: This is a multi-turn code repair benchmark where the model must repeatedly adjust code based on compiler errors, runtime exceptions, and other feedback. The model scored 51.7%, far exceeding Overseas Flagship A (42.5%) and Flagship B (46.3%), demonstrating strong context retention and iterative error-correction capabilities. This is particularly important for handling long-standing bugs in real projects.
• Agents’ Last Exam: This benchmark comprehensively evaluates the model’s ability to act as an agent, including planning, tool calling, and reasoning. The model led with a score of 43.2, 10.4 points higher than Overseas Flagship A.

A clear weakness: 3D geometric understanding remains a challenge.

BenchCAD tests the accuracy of reconstructing 3D scenes from text descriptions or 2D images. The model scored 92.7%. While the absolute number is close to a perfect score, it still trails Overseas Flagship A (95.9%) and Flagship B (96.2%) by more than 3 percentage points. This suggests that developers working on high-precision geometric tasks such as 3D spatial reasoning, CAD-assisted design, or robot path planning may need to consider using specialized models or adding result verification.

03 Pricing Strategy: The Logic Behind One-Fifth the Cost and Selection Guidance

The most striking part of this release is that its pricing is only one-fifth that of a comparable overseas flagship model. This is not simply a price war; it is based on the natural cost advantages of the MoE architecture:

1. Low activated parameters: The effective parameter count in the hundreds-of-billions range significantly reduces the compute resources required for each inference. Cloud providers can therefore offer API services at a much lower marginal cost.
2. Long-context friendliness: The 1 million token context window reduces the engineering complexity of splitting multi-turn sessions and building retrieval-based QA systems, saving on additional infrastructure such as vector databases and reranking services.
3. Optimized batch throughput: Efficient expert routing makes it easier to schedule high-concurrency requests on the server side, further reducing the cost per call.

Developer selection guidance:

• Scenario 1: Code generation and rapid iteration. For small and medium-sized teams that need frequent API calls and are sensitive to token consumption, such as independent developers and startups, this model can serve as the primary coding assistant, covering high-frequency tasks including code explanation, unit test generation, and README writing.
• Scenario 2: Large-scale document analysis. The model can ingest up to 1 million tokens of documents at once, such as annual financial reports, technical white papers, and legal contracts, for summarization and question answering. This can significantly reduce manual review time without requiring a complex in-house RAG pipeline.
• Scenario 3: High-concurrency agent applications. For building agent services such as automated customer support and data collection and organization, low cost plus strong performance means lower marginal operating costs and a more favorable business model calculation.

04 Integration Practices and General Performance Tuning Guide

For technical teams, the key practical challenge is how to integrate such an MoE large model into existing business workflows quickly and cost-effectively. The following recommendations are based on general best practices and do not depend on any specific cloud provider:

1. Context compression strategies
Although the context window supports up to 1 million tokens, extremely long contexts still affect time-to-first-token latency. For structured logs or code files, consider syntax-level compression first, such as removing comments and blank lines, rather than directly concatenating all content.

2. Temperature and Top-P tuning
For code generation tasks, set Temperature between 0.1 and 0.3 to maintain deterministic output. For creative copywriting or brainstorming tasks, raise it to 0.7 to 0.9 to encourage diversity.

3. Tool calling structure design
The model natively supports tool calling, also known as function calling. When designing agent workflows, prioritize structured JSON Schema descriptions for tool inputs and outputs, and avoid natural language-based intermediate parsing.

4. Caching for high-frequency scenarios
For highly repetitive prompt prefixes, such as system role definitions and fixed-format templates, implement prefix caching at the application layer to reduce latency and cost from repeated computation.

05 Summary and Outlook

Chinese large models are undergoing a critical shift from parameter competition to engineering cost-effectiveness. This flagship product demonstrates with its 3-trillion-parameter MoE architecture that extreme model capacity and controlled inference costs can coexist. While it still shows small gaps against top-tier closed-source models in areas such as 3D spatial understanding, it has achieved practical substitution capability in core productivity domains including coding agents and long-context processing. For developers seeking controllable costs and reliable task performance, this model is currently a strong option worth evaluating. As more blind benchmark data becomes public and community feedback accumulates, future iterations may address remaining weaknesses and further strengthen its position as a cost-effective flagship.