As of September 8, 2026, Moonshot has reported on large-scale training with the Muon optimizer and released a distributed implementation and several types of model checkpoints; Zhipu and Tencent showcase models and products for Agents, long-running tasks, or on-device deployment on their respective pages; an AWS tutorial deploys communication channels, ordering logic, and shared memory in separate layers.[S2][S5][S6][S8] The KOPA-Bench study uses live APIs to validate connections between tools, synthesizes multistep trajectories, and trains models using the resulting data and GRPO. The paper's abstract reports that the trained 9B model approaches the untuned 27B model from the same family, but does not separately measure the contributions of live validation, data synthesis, and GRPO.[S10] ROBORMBENCH finds that, in the robotic reward tasks tested, neither scaling up models nor adding explicit reasoning reliably eliminates instability caused by paraphrasing.[S9] These materials indicate that model training, task adaptation, state management, and execution validation can be examined as separate stages. If enterprises continue to place more task state and tool connections in Agent runtime platforms, competition may expand to encompass the entire execution chain. However, vendor releases, deployment tutorials, and domain-specific benchmarks do not establish that the global market has already shifted from foundation models to runtimes, nor can open models or checkpoints be directly equated with lower costs or reduced platform lock-in.
Beyond Models, Agent Systems Also Handle Channels, Memory, and Tool Execution
Foundation models can be called directly by applications or trained further for specific tasks. In the Agent scenarios presented in these materials, evaluation also encompasses state persistence across channels, multistep tool calls, and the stability of execution results. The AWS example demonstrates a channel layer, an ordering layer, and shared memory; KOPA-Bench and ROBORMBENCH test multistep tool calls and robotic reward judgments, respectively.[S2][S9][S10]
The WhatsApp ordering assistant demonstrated by AWS can receive orders through text, voice messages, and live voice calls. The example separates communication channels from ordering logic and uses shared memory to recognize the same customer across all three channels.[S2] Zhipu's page lists foundation models, Agent-related models, application-level APIs, fine-tuning, and search tools; Tencent's page showcases different offerings, including a research Agent, an office-work delivery pipeline, foundation models, and on-device translation.[S5][S8]
These projects show that models can be deployed alongside channels, memory, tools, and application layers, but they do not establish that models have become standard components that can be replaced at no cost. The materials provide no data on customer migrations, switching time, or switching costs, nor do they include standardized testing of interface compatibility across projects.
If task histories and tool connections are concentrated in a single runtime platform, whether migration requires additional modifications and validation will depend on the platform's interfaces, its data export capabilities, and the target system's compatibility. The AWS example demonstrates only one layered architecture; it does not measure migration difficulty or indicate this architecture's share of the market.[S2]
Nine Technical Developments Span Training, Execution, and Validation
The following nine items are an editorial classification of the materials by technical stage, not industry statistics. Performance descriptions on vendor pages are vendor-reported claims; academic benchmarks can reveal capabilities or failure modes in specific tasks, but cannot be translated directly into commercial adoption rates.
| Stage | Observable Facts | Potential Impact | Limits of Applicability |
|---|---|---|---|
| Training efficiency | Moonshot reports that, after adding weight decay and adjusting the parameter update ratio, Muon achieved approximately twice the computational efficiency of AdamW in the relevant scaling experiments; the team also released a distributed implementation, model checkpoints, and intermediate checkpoints.[S6] | If other teams reproduce this result under comparable settings, less computation may be needed to achieve comparable training performance. | The result comes from specific models and training settings reported by Moonshot and cannot be directly extrapolated to training costs for all models. |
| Training dynamics | Tencent research introduces the effective learning rate, ELR, to control changes in the direction of model weights, and reports that the optimal ELR transfers more reliably across model scales.[S8] | If this result can be reproduced across models and institutions, some cross-scale hyperparameter-tuning experiments may be reduced. | The information comes from a research summary on Tencent's page, contains no commercial cost data, and is no substitute for external replication. |
| Open training artifacts | Moonlight is a mixture-of-experts model with 3B/16B parameters, trained on 5.7 trillion tokens. Moonshot says it has released pretrained, instruction-tuned, and intermediate checkpoints, along with a distributed Muon implementation.[S6] | These releases allow external teams to access and inspect some models and training artifacts. | The materials do not fully specify licensing terms, the extent of training-data openness, or the cost of reproducing the entire process. |
| Specialized Agent models | Zhipu's page says GLM-5-Turbo optimizes tool calling and long-chain execution for Agent scenarios, and lists Agent products such as AutoGLM.[S5] | The page shows that Zhipu is providing specialized models or products for tool use and sequential tasks. | The descriptions come from the vendor's general homepage, with no standardized external success-rate data, cost comparisons, or long-term stability data. |
| Cloud Agents | The AWS tutorial uses Amazon Bedrock AgentCore and Amazon Nova 2 to deploy a multimodal ordering assistant, separating the channel layer, ordering layer, and shared memory.[S2] | The example demonstrates one way to implement state management, channel adaptation, and ordering logic in separate layers. | The tutorial does not establish the scale of adoption, effects on customer retention, or the distribution of commercial value. |
| Tool calling | KOPA-Bench contains 145 real-world tasks. The study uses live APIs to validate connections between tools, synthesizes multistep trajectories, and then trains a 9B model using the resulting data and GRPO; the paper's abstract says the trained model approaches the untuned 27B model from the same family.[S10] | Within this task setting and model family, the combined process reported in the study corresponds to a smaller gap between the 9B model and the 27B model from the same family. | The abstract does not provide specific scores here or separately identify the contributions of live validation, synthetic data, and GRPO; the result cannot be generalized to all tool-calling tasks. |
| On-device models | Tencent's page says Hy-MT2 supports 33 languages; its 1.8B version requires 440MB of storage after 1.25-bit quantization and can perform local inference on the various mobile chipsets listed on the page.[S8] | If speed, energy consumption, and quality on actual devices meet application requirements, some translation tasks can use local inference. | Storage size and local inference capability alone do not establish fewer cloud calls, lower costs, improved privacy, or user adoption; performance comparisons are also vendor-reported claims. |
| Speech specialization | MiniMax says Speech 2.8 introduces native support for fillers for the first time, along with natural tone tags; the page also says the model can clone a voice from a 10-second original voice clip.[S7] | The product offers conversational fillers, pauses, and voice cloning as specific speech-interaction features. | The page primarily consists of product descriptions and demonstrations and cannot serve as an independent quality evaluation. |
| Reliability | ROBORMBENCH contains 2390 real robot trajectories and 21673 validated paraphrases. The study reports that paraphrasing a goal description can flip the judgment of the same behavior between failure and success, and that neither scaling up models nor adding explicit reasoning reliably eliminates this problem.[S9] | In similar robotic reward tasks, model stability under meaning-preserving paraphrases can be tested separately. | This conclusion cannot be directly generalized to all enterprise Agent workflows. |
Training-Efficiency Improvements Emerge Alongside Larger Models
Moonshot describes Moonlight as a mixture-of-experts model with 3B/16B parameters, trained on 5.7 trillion tokens.[S6] The materials do not further explain how the model selects experts during each inference, so only its parameter scale, architecture name, and training-data volume can be confirmed here. Moonshot also reports that, after weight decay and parameter update ratio adjustments, Muon achieved approximately twice the computational efficiency of AdamW in its scaling experiments.[S6]
Tencent's ELR research addresses control of training dynamics. The page says a conventional learning rate cannot precisely control how quickly model parameters evolve, and its preferred values also change with model and data scale; ELR instead focuses on changes in weight direction and is used to predict loss and design learning-rate schedules.[S8] These are research findings summarized on Tencent's page and cannot be translated directly into commercial training expenses.
If methods such as Muon or ELR repeatedly achieve similar effects across different teams, model scales, and data settings, the experimentation or computational investment required to reach comparable training performance may decrease. Further inferences about an increase in developer numbers, lower API prices, or changes in revenue structure would require pricing, budget, adoption, and revenue data; the current materials contain none of these metrics.
Meanwhile, Tencent's page says Hy3 uses a MoE architecture with 295B total parameters and 21B active parameters, and supports a 256K context window.[S8] A more cautious conclusion is that these materials show two concurrent paths—scaling up models and improving training methods—without yet establishing which will dominate the market.
Agent Examples Bring Shared Memory and Tool Execution into the Runtime Layer
The assistant in the AWS example receives orders through text, voice messages, and live voice calls, and uses shared memory to recognize the same customer across all three channels.[S2] KOPA-Bench requires models to call open public APIs across multiple steps.[S10] These two sources demonstrate cross-channel state and multistep tool execution, respectively, but do not provide a universal architecture applicable to all Agent systems.
The AWS example separates the channel layer from the ordering layer and uses Amazon Nova 2 and shared memory on Amazon Bedrock AgentCore.[S2] This shows that cloud components can handle channel and state operations beyond the model itself, but a deployment tutorial does not establish that the commercial value of cloud services has shifted from compute to state management.
If an actual deployment stores task histories and tool connections in a particular platform, the migration workload will depend on whether these can be exported, whether interfaces are compatible, and whether revalidation is needed. The current materials contain no migration cases or switching-cost measurements, so this possibility cannot be described as platform lock-in that has already occurred.
The KOPA-Bench study builds a graph of connections between one tool's output and another tool's input, retains only connections that succeed when tested with live API calls, and uses them to synthesize multistep trajectories; it then trains models using the resulting data and GRPO.[S10] The paper's abstract reports that the trained 9B model approaches the untuned 27B model from the same family.[S10] This result follows the combined use of live validation, trajectory synthesis, and GRPO. The abstract does not separately measure each stage's contribution, and the narrowing of the gap cannot be attributed to any one stage.
The Qwen3Guard article listed on the Qwen blog index page was published on September 23, 2025. The index summary describes it as the Qwen family's first safety guardrail model, capable of classifying prompts and responses and outputting risk levels and categories across Chinese, English, and multilingual settings.[S3] This information indicates that Qwen3Guard is intended for safety classification of prompts and responses, but the index summary does not show specific deployment configurations, its position in Agent workflows, or incident rates, and therefore does not establish that execution risks are under control.
Open Models and Training Artifacts Increase Accessibility, but Costs and Migration Effects Require Separate Assessment
Moonshot explicitly states that it has released the distributed Muon implementation, along with Moonlight's pretrained, instruction-tuned, and intermediate checkpoints.[S6] Zhipu's page describes GLM-5.2 as a flagship model that is now available and open-source.[S5] These are openness initiatives by two specific projects and cannot be used to generalize about the pace or scope of openness among all Chinese model vendors.
Access to models, code, or checkpoints gives external teams an opportunity to inspect and try them. The materials do not fully provide these projects' licensing terms, the extent of training-data openness, hardware requirements, or standardized compatibility tests, so they cannot be used to quantify deployment costs or replacement difficulty.
If the model files, task state, and tool connections required by an application can be migrated separately, replacing the model may involve fewer modifications; if some of these cannot be migrated directly, additional work may be needed. This architectural judgment still requires testing through actual migrations and cannot be treated as a measured market outcome.
Similar limitations apply to on-device deployment. Tencent's page says the quantized 1.8B version of Hy-MT2 requires 440MB of storage and can perform local inference on the mobile chipsets listed on the page.[S8] The relevant tasks may use local inference only if speed, energy consumption, output quality, and device coverage meet practical requirements; 440MB of storage and the ability to run locally do not by themselves establish fewer cloud calls, lower expenses, or improved privacy.
How to Test These Technical Paths Next
- As of September 8, 2026, what can be confirmed is that Moonshot, Zhipu, Tencent, and AWS have each demonstrated training methods, foundation models, Agent-related capabilities, or layered runtime examples.[S2][S5][S6][S8] Within the tool-calling tasks defined by KOPA-Bench, the study's use of live validation, trajectory synthesis, and GRPO training brought a 9B model close to the untuned 27B model from the same family, but the abstract does not separately measure each stage's contribution.[S10] Within the robotic reward tasks defined by ROBORMBENCH, neither scaling up models nor adding explicit reasoning reliably eliminates paraphrase fragility.[S9] These results support evaluating training, execution, and validation separately, but do not establish that global competition, revenue, or market share has shifted from models to runtimes.
- On the training side, further observation can assess whether Muon and ELR can be reproduced across institutions, model scales, and data settings, and compare the computation, training time, and total cost required to achieve comparable results.[S6][S8] Until replicated, the approximately twofold computational efficiency reported by Moonshot cannot be treated as a universal effect.
- On the Agent side, useful metrics to track include real-world task completion rates, tool-call error rates, human takeover rates, state recovery performance, and actual migration time. The current materials do not provide these commercial deployment metrics, so they cannot establish whether layered runtime architectures have already delivered reliability or cost advantages.
- For the open ecosystem, licensing, weight and code availability, the extent of data disclosure, hardware requirements, and interface compatibility can be documented separately. Releasing models or checkpoints does not by itself establish cheaper deployment or reduced platform lock-in.[S5][S6]
- Reliability evaluations need to maintain scenario-specific boundaries. WearableQA contains 4084 questions with ten answer options each, drawn from wearable time series, blood biomarkers, and demographic information from 200 real users, with up to 500 days of daily measurements per user and 16 question categories.[S1] This material shows that a dedicated benchmark exists for real-world longitudinal health data, but the abstract provides no test results that could be used to assess the effects of model scale.
- The findings on paraphrase stability in robotic reward models can suggest similar tests for other Agents, but cannot replace validation of enterprise workflows themselves.[S9] Likewise, the AWS tutorial, the general Zhipu and Tencent pages, and MiniMax's product demonstrations are suitable for illustrating specific designs and vendor claims, but cannot substitute for adoption rates, costs, migration difficulty, and independent quality evaluations.[S2][S5][S7][S8]
Keep following engineering practices
Subscribe to new articles with an RSS reader, or share this article with someone working on similar problems.
Subscribe to RSS updates