Open-Weight AI Models Enter a New Phase: A Technical Briefing on Ultra-Large Mixture-of-Experts Models

    This article reviews the release of weights for an ultra-large mixture-of-experts model with about 2.8 trillion total parameters and roughly 104 billion activated parameters per inference. It explains why MoE architectures attract attention, why total parameters and activated parameters should not be confused, how to interpret evaluation rankings, what open weights provide in practice, and what enterprises should assess when deploying and selecting ultra-large models.

    Open-Weight AI Models Enter a New Phase

    1. Today’s Briefing Overview

    A large-model team in China has officially released the weight files for an ultra-large mixture-of-experts model according to a previously announced schedule. The model has about 2.8 trillion total parameters and activates roughly 104 billion parameters per inference. It is among the larger representative open-weight models currently available.

    Public information shows that the model achieved a high ranking in open-model evaluations for AI agents and performed well on a comprehensive model evaluation leaderboard in China. These results suggest that open-weight models are moving from “can run” toward “can handle complex tasks,” especially in long-context understanding, tool calling, task planning, and multi-step execution. This points to stronger potential for engineering applications.

    However, evaluation rankings do not equal absolute performance in production. When adopting such models, enterprises still need to assess them against actual business data, inference cost, deployment conditions, response stability, and security requirements.

    2. Why Mixture-of-Experts Architecture Attracts Attention

    The core idea of a mixture-of-experts architecture is to activate only part of the model’s parameters for different inputs, rather than using all parameters in every computation.

    A typical MoE model usually contains multiple expert networks and a router module. The router selects a number of experts that are better suited to the current task based on the input. This allows the model to have a very large total parameter capacity while controlling the computation required for a single inference.

    This can be simplified as:

    1. Input
    2. ↓
    3. Router
    4. ↓
    5. Select a small number of expert networks
    6. ↓
    7. Expert computation and result fusion
    8. ↓
    9. Output

    This design provides three benefits.

    First, model capacity is larger. Different experts can gradually develop relative specialization during training. For example, some experts may become better at code understanding, some at natural language expression, and others at mathematical reasoning or knowledge question answering. Although this division is not fully explicit, large-scale training can often produce a degree of capability specialization.

    Second, inference computation is more flexible. A model may have trillions of total parameters, but only a subset is activated for each request. The theoretical computation cost can therefore be lower than that of a dense model with the same total parameter scale.

    Third, the scaling path is clearer. Research teams can continue improving model capacity and task coverage by increasing the number of experts, optimizing routing strategies, and improving expert load balancing.

    However, MoE architectures also face clear challenges. If routing leads to uneven load, a small number of experts may be called frequently while others are underutilized. This can affect training efficiency and may create inference bottlenecks. In addition, the very large total parameter scale makes weight storage, VRAM allocation, parallel communication, and fault recovery more complex.

    3. Total Parameters and Activated Parameters Are Not the Same

    One of the easiest misunderstandings when reading large-model news is to treat total parameter count as directly equivalent to the computation required for a single inference.

    Total parameters represent the scale of all trainable parameters in the model and reflect overall model capacity. Activated parameters represent the parameter scale actually involved in computation after routing for one input, which is closer to the computational load of a single inference.

    For example, a model may have trillions of total parameters but activate only about 100 billion parameters each time. Its knowledge capacity and expressive capability may approach those of an ultra-large model, but its hardware requirements will not simply equal those of a dense model that uses all parameters on every run.

    It is important to note that fewer activated parameters do not necessarily mean deployment cost will be low. Actual cost is also affected by the following factors:

    1. Whether weights need to reside fully in memory or VRAM.
    2. Whether experts require cross-device communication.
    3. Whether routing results cause uneven distribution of computation.
    4. Whether context length and concurrent request volume are high.
    5. What precision is used for inference.
    6. Whether quantization, caching, and batching techniques are used.
    7. How well the inference framework supports MoE structures.

    Therefore, enterprises cannot determine whether a model is suitable for deployment solely from activated parameters. Conclusions should be based on real business stress testing.

    4. How to Interpret Evaluation Rankings

    The model’s performance on the open-weight leaderboard and in a domestic comprehensive evaluation reflects broad recognition of its overall capabilities. However, evaluation results usually only show performance on specific datasets, with specific prompts and specific scoring rules. They cannot directly replace business validation.

    From a technical perspective, model evaluations can be roughly divided into four categories.

    The first is general knowledge and language understanding evaluation, used to observe a model’s information organization, reading comprehension, factual answering, and language expression.

    The second is reasoning evaluation, focusing on mathematical problems, logical deduction, complex conditional analysis, and multi-step problem solving.

    The third is code and tool-use evaluation, used to judge whether a model can generate executable code, understand interface descriptions, call external tools, and adjust its approach based on execution results.

    The fourth is AI agent evaluation, which focuses on whether a model can decompose goals, make plans, maintain context, handle exceptions, and complete multi-turn tasks.

    For enterprise applications, the fourth category is becoming increasingly important. Real business workflows are often not simple question-and-answer interactions. They typically involve requirements analysis, information retrieval, rule judgment, tool calling, result validation, and final delivery.

    However, agent evaluations also need to be interpreted carefully. Strong performance on standard tasks does not mean a model can directly take over enterprise processes. Production environments also require permission control, data isolation, operation auditing, error rollback, and human confirmation mechanisms.

    5. The Practical Value of Open Weights

    The significance of open weights is not just that a model can be downloaded. More importantly, it lies in verifiability, adaptability, and controllability.

    First, verifiability. Developers can test the model locally under the applicable license terms and observe output stability, resource consumption, and business fit, rather than relying entirely on documentation for a remote API.

    Second, adaptability. Enterprises can optimize prompts, add retrieval augmentation, perform supervised fine-tuning, or use parameter-efficient fine-tuning around their own data so that the model better fits internal knowledge, industry terminology, and process rules.

    Third, controllability. For scenarios involving sensitive data, internal documents, or high compliance requirements, localized deployment or dedicated-environment deployment can reduce the risk of data leaving the internal boundary. Of course, open weights do not automatically mean security and controllability. Enterprises still need a complete security governance system.

    In practice, open-weight models are often suitable for the following scenarios:

    • Internal enterprise knowledge Q&A.
    • Code assistance and R&D document generation.
    • Structured information extraction.
    • Multi-step process automation.
    • Private data analysis.
    • Industry-specific AI agents.
    • Business systems with high requirements for data residency and access auditing.

    6. What to Focus on When Deploying Ultra-Large Models

    For ultra-large MoE models, deployment should not focus only on whether the model can start. A complete engineering evaluation process should be established.

    The first step is to confirm hardware resources. This requires assessing model weight size, inference precision, VRAM capacity, host memory, device interconnect bandwidth, and concurrency targets. Even with low-precision weights, the model may still need to run across multiple devices.

    The second step is to choose an appropriate parallelism approach. Common approaches include data parallelism, tensor parallelism, pipeline parallelism, and expert parallelism. Different approaches suit different model structures and hardware environments. MoE models especially require attention to expert distribution and cross-device communication; otherwise, computation may wait too long.

    The third step is to validate quantization. Quantization can reduce storage and inference costs, but it may affect complex reasoning, code generation, and long-text stability. It is not enough to look only at average scores. Key business samples should also be tested, especially tasks involving numbers, rules, and format constraints.

    The fourth step is to establish stress-testing metrics. At a minimum, teams should record time to first token, average generation speed, concurrent throughput, VRAM usage, request failure rate, long-context degradation, and continuous-operation stability.

    The fifth step is to design fault-handling mechanisms. Production systems should support request timeouts, service degradation, model switching, result retry, and human takeover, so that an anomaly in a single model does not affect the entire business chain.

    7. How Enterprises Can Build Their Own Model Selection Criteria

    Faced with a growing number of open models, enterprises can use a three-dimensional evaluation method covering capability, cost, and governance.

    At the capability level, the focus should be on whether the model actually solves business problems, not just on public leaderboard rankings. Teams should prepare a set of de-identified real samples covering common tasks, boundary tasks, and failure tasks, and test them with consistent prompts and scoring rules.

    At the cost level, in addition to hardware purchase or rental costs, enterprises should calculate hidden costs such as model loading time, operations labor, data storage, network communication, log retention, and security auditing.

    At the governance level, organizations need to check license scope, training-data disclosure, content safety capabilities, output explainability, sensitive-information handling, and version management mechanisms. For high-requirement fields such as finance, healthcare, and government affairs, stricter approval and audit processes should also be established.

    General-purpose capabilities from cloud platforms such as Baidu AI Cloud—including compute, containers, object storage, logging, monitoring, and security—can be used to build infrastructure for model deployment and operations. However, specific selection should still be based on the enterprise’s compliance requirements, data boundaries, and actual workload. Cloud resource capabilities should not be equated simply with the model’s own capabilities.

    8. Trend Judgment

    The release of this ultra-large open-weight model sends three notable signals.

    First, competition among open models is shifting from pure parameter scale toward comprehensive engineering capability. Model architecture, training efficiency, inference optimization, tool use, and the deployment ecosystem will jointly determine practical value.

    Second, AI agents will become an important application direction. Future models will not only answer questions but also understand goals, call tools, execute workflows, and self-check results. Evaluation systems will therefore gradually shift from static question answering to dynamic task completion rate.

    Third, model selection will become more scenario-specific. Enterprises do not necessarily need the model with the largest parameter count. They need a balance among effectiveness, cost, speed, privacy, and maintainability. For some tasks, medium-sized models combined with retrieval systems, rule engines, and workflow orchestration may be more stable than having an ultra-large model directly handle all tasks.

    Overall, the rapid development of open-weight models gives developers and enterprises more technical choices, while also raising the bar for model evaluation and systems engineering. The right practice is not to chase leaderboards blindly, but to establish a complete loop from sample preparation and offline evaluation to resource stress testing and online monitoring. Only when model capabilities can be reliably converted into controllable, maintainable, and auditable business workflows can the technical value of open weights truly be realized.