{"@context": "https://schema.org", "@type": "AnalysisNewsArticle", "inLanguage": "en", "identifier": "ai-research-20260906-0025-766", "headline": "Moonshot Says Muon Is About Twice as Compute-Efficient as AdamW for Training; Zhipu and Qwen Showcase Agent and Safety Models", "description": "Several preprints also examine reusable local neural functions, auxiliary representations of knowledge, and shorter optimized prompts.[S1][S8][S10] Whether these methods will change the mix of cloud-based calls depends on reproducibility across settings, the scale of production adoption, and total costs.", "abstract": "This article draws on nine substantive sources, including papers, technical articles, announcements, and product pages. The publication dates shown on vendor materials range from March 2025 to September 2026; S1, S8, S9, and S10 are arXiv records updated on September 3, 2026, while Zhipu's official website and the Qwen blog index are treated as snapshots captured on September 6, 2026.[S1][S2][S3][S5][S6][S7][S8][S9][S10] Moonshot says that, under its compute-optimal training conditions, Muon achieved approximately twice the compute efficiency of AdamW; the project also released a distributed implementation and model checkpoints from multiple training stages.[S6] Four preprints examine auxiliary representations of knowledge, reusable local neural functions, prompt optimization, and the assessment of reasoning-step importance, respectively.[S1][S8][S9][S10] On the product side, the snapshot of Zhipu's official website showcases open-source models, agent foundation models, and hosted services; the Qwen blog index includes an article on a safety classification model dated September 23, 2025; MiniMax's page lists January 23, 2026, as the release date for Speech 2.8; and OpenAI announced a $1 billion commitment to cybersecurity AI for essential services on September 3, 2026.[S2][S3][S5][S7] These results come from specific experiments or vendor releases and do not, on their own, establish that development workflows, market shares, or cloud workloads have broadly changed. If the methods can be reproduced across models, hardware, and business settings, and cover enough frequent, stable tasks, some applications may reduce their use of large remote models on a per-call basis; complex, rapidly changing tasks or those requiring centralized management may continue to use cloud services.", "datePublished": "2026-09-06T00:25:46+00:00", "dateModified": "2026-09-25T14:22:46.766747+00:00", "author": {"@type": "Person", "name": "Aperture Research Editorial Team", "jobTitle": "Industrial Organization and Technological Change Research · Research Coder"}, "about": ["Foundation models", "AI Agent", "Open-source models", "Cloud AI infrastructure", "Local inference", "Training efficiency", "Safety guardrails", "Large-model products"], "category": "AI industry trends", "authorityScore": 98, "claims": [{"text": "The nine substantive sources used in this article include papers, technical articles, announcements, and product pages. The publication dates shown on vendor materials range from March 2025 to September 2026; all four arXiv records show updates on September 3, 2026, while Zhipu's official website and the Qwen blog index are treated as snapshots captured on September 6, 2026.[S1][S2][S3][S5][S6][S7][S8][S9][S10] On the training side, the sources cover the Muon optimizer and experiments with auxiliary representations of knowledge; on the execution side, they cover neural functions that can be saved and versioned, as well as shorter automatically optimized prompts; on the product side, they include agent foundation models, safety classification, voice interaction, and cybersecurity support for essential services. Another study tests the ability of LLM judges to identify important reasoning steps.", "section": "New Developments in Training, Execution, and Safety Components; Industry Impact Depends on Reproducibility and Adoption"}, {"text": "These materials come from different organizations and evaluation environments. The preprints report specific benchmarks or controlled experiments, while vendor webpages provide product announcements or vendors' own descriptions of their capabilities. In particular, S3 is a captured snapshot of the Qwen blog index, and S5 is a captured snapshot of Zhipu's dynamic official website.[S3][S5] They can establish which products and technical directions the pages described at the time, but they cannot establish that the global market has broadly shifted from remote model calls to local execution, nor can the components showcased by different vendors be treated as an already integrated, unified agent product stack.", "section": "New Developments in Training, Execution, and Safety Components; Industry Impact Depends on Reproducibility and Adoption"}, {"text": "If training-efficiency methods can be reproduced reliably and compilable functions can cover enough real-world tasks, application teams may redistribute work among general-purpose models, compact models, and local modules. The ultimate impact on cloud services will also depend on the balance between simple tasks moving away from the cloud and new complex tasks being added; these sources provide no data on workloads, the scale of adoption, or total cost of ownership.", "section": "New Developments in Training, Execution, and Safety Components; Industry Impact Depends on Reproducibility and Adoption"}, {"text": "The table below treats paper-based experiments separately from product vendors' releases. Performance-related statements retain the framing used by their sources, without turning individual benchmarks, demonstrations, or vendor rankings into industry-wide conclusions.", "section": "What the Nine Sources Establish, from Training to Interaction"}, {"text": "These materials provide no figures for global cloud market share, enterprise agent adoption, API revenue, compute procurement volumes, or production incident rates. The industry implications in the table are therefore conditional analyses, not evidence that market outcomes have already occurred.", "section": "What the Nine Sources Establish, from Training to Interaction"}, {"text": "The Muon project focuses its design adjustments for large-scale training on two areas: incorporating weight decay and more precisely adjusting the relative update scales for different parameters. The project team says these adjustments improve training stability; in scaling experiments using the complete approach under compute-optimal training conditions, Muon was approximately 2 times as compute-efficient as AdamW.[S6] The material does not separately quantify each adjustment's contribution, so the overall efficiency result cannot be attributed to either one alone or directly extrapolated to all models, hardware, and training configurations.", "section": "Training Efficiency Is Influenced by Both Optimization Algorithms and Data Representation"}, {"text": "The project describes Moonlight as a 3B/16B-parameter mixture-of-experts model trained with Muon on 5.7 trillion tokens, and it released code, pretrained and instruction-tuned checkpoints, and intermediate checkpoints.[S6] These assets give external teams access to the project's implementation and models from different training stages, but actual replication still requires the corresponding compute, data processing, and engineering resources.", "section": "Training Efficiency Is Influenced by Both Optimization Algorithms and Data Representation"}, {"text": "The knowledge-learning experiments offer another route. Under a fixed token budget, the researchers replaced some tokens used to repeat documents with auxiliary representations of the same knowledge and observed improved learning in controlled experiments.[S10] This shows that, under the conditions defined in the paper, the way knowledge is represented affects learning outcomes, but it does not directly imply equivalent gains in commercial pretraining pipelines. The date shown for S10 is the arXiv update date, not the paper's initial publication date.", "section": "Training Efficiency Is Influenced by Both Optimization Algorithms and Data Representation"}, {"text": "If Muon's results can be reproduced across model scales, hardware, and data distributions, the same compute budget might support more training attempts or a larger individual training run; how that budget is allocated would still depend on a team's resources and goals. If the gains depend on a specific architecture or training configuration, the direct role of releasing code and checkpoints is primarily to make the implementation and model assets more accessible, rather than to lower training costs universally.", "section": "Training Efficiency Is Influenced by Both Optimization Algorithms and Data Representation"}, {"text": "Compile by Training addresses a specific class of problems: tasks that are easy to describe in natural language but difficult to encode as handwritten rules, where calling a large remote model each time incurs recurring costs, latency, and vendor dependence. The method calls a teacher model during compilation to generate task examples and trains small adapters for a compact interpreter; the teacher model is no longer required at runtime, but local inference still has to be performed.[S1]", "section": "Compilable Functions May Reduce Some Repeated Calls, Provided Tasks Are Stable Enough"}, {"text": "This arrangement concentrates teacher-model calls and task-adaptation costs in the compilation stage rather than eliminating all runtime costs. In the settings reported by the paper, compilation takes about one minute, and the resulting functions can be saved, versioned, and composed.[S1] If a task is called frequently, its input boundaries are stable, and the combined costs of local inference, recompilation, and maintenance are manageable, this approach may be more suitable than making a remote call each time; infrequent or frequently changing tasks may not be suitable.", "section": "Compilable Functions May Reduce Some Repeated Calls, Provided Tasks Are Stable Enough"}, {"text": "ESPO first diagnoses error structure, then generates candidate prompts using multiple strategies, and finally performs stability selection. Across seven benchmarks, it exceeds GEPA by an average of 3.76 percentage points while using prompts that are 47% shorter.[S8] Both results occurred together in these experiments, but this does not establish that shorter prompts caused the accuracy improvement. It shows only that, on these benchmarks, achieving higher average accuracy does not require continually adding rules and exceptions.", "section": "Compilable Functions May Reduce Some Repeated Calls, Provided Tasks Are Stable Enough"}, {"text": "These experiments do not establish that remote-interface call volumes or revenue have already been affected. Some remote calls may decline only if compilable methods cover a large number of production tasks and the combined costs of local inference, deployment, updates, monitoring, and error handling are lower than those of remote calls. The sources above provide no figures for the share of tasks migrated, total cost of ownership, or changes in revenue.", "section": "Compilable Functions May Reduce Some Repeated Calls, Provided Tasks Are Stable Enough"}, {"text": "Different vendors showcase capabilities at different stages: Moonshot released a training method, code, and checkpoints; the snapshot of Zhipu's official website showcases open-source models, tool calling, and multimodal agent foundation models; the article listed on the Qwen blog index introduces a model for input and output safety classification; and MiniMax's page introduces a speech model with tone control and voice cloning.[S3][S5][S6][S7]", "section": "Different Vendors Showcase Different Components, but They Cannot Be Treated as a Single Product Stack"}, {"text": "These releases and webpages show that different companies separately offer or introduce training, execution, safety, and interaction components. The materials do not show that the components already interoperate, use common standards, or have been jointly deployed, so they cannot be treated as a single product stack or used to judge how quickly the supply or market adoption of these products is changing.", "section": "Different Vendors Showcase Different Components, but They Cannot Be Treated as a Single Product Stack"}, {"text": "At the time of capture, Zhipu's official website also displayed open-source models, API services, model fine-tuning, AI search tools, and application-level interfaces side by side.[S5] This establishes only that the dynamic page presented open models and hosted services together at that time. It does not prove that open models will necessarily weaken its API business, nor that commercial services will become its main source of revenue. The snapshot also cannot confirm the relevant GLM models' individual release dates, prior status, continued subsequent availability, or whether their rankings can be sustained over time.", "section": "Different Vendors Showcase Different Components, but They Cannot Be Treated as a Single Product Stack"}, {"text": "If enterprises can migrate smoothly between models, tools, and deployment environments, open assets may expand their options. If migration remains constrained by format, data, engineering, and operational requirements, the original vendor's hosted services may still offer convenience. The materials above contain no data on migration costs, deployment choices, or revenue, so they cannot establish which of these effects is stronger.", "section": "Different Vendors Showcase Different Components, but They Cannot Be Treated as a Single Product Stack"}, {"text": "The article listed in the captured Qwen blog index describes Qwen3Guard as capable of classifying both prompts and responses and outputting risk levels and categories.[S3] Under this functional definition, if a system integrates the model into its generation workflow, it can be used to classify the textual safety of inputs and outputs; the index does not explain how specific systems integrate it or provide deployment-scale figures.", "section": "Safety Models Can Handle Classification, but Governance Outcomes Depend on Real-World Operation"}, {"text": "Whether safety classification can become a reliable infrastructure component depends on whether it can be continuously integrated into production systems and operate with acceptable false-positive rates, false-negative rates, and latency. The available material provides no adoption rates, changes in incidents, or actual operating costs, so it can confirm only the functional direction introduced by the article listed on the index, not governance outcomes.", "section": "Safety Models Can Handle Classification, but Governance Outcomes Depend on Real-World Operation"}, {"text": "OpenAI committed $1 billion to cybersecurity AI, training, and support for essential services.[S2] The commitment indicates a spending plan, but the material does not state the progress of fund disbursement, the number of participating organizations, or security outcomes, and it cannot establish a change in market power.", "section": "Safety Models Can Handle Classification, but Governance Outcomes Depend on Real-World Operation"}, {"text": "When reasoning traces are used for auditing, it is also necessary to distinguish whether the text is readable from whether a judge model can identify the steps that genuinely affect the outcome. The experiments in S9 show that, in its step-importance identification task, stronger LLM judges outperform the prevalence baseline but remain well below the noise ceiling.[S9] Thus, under similar experimental settings, judging which steps are more important solely from chain-of-thought text may not be sufficiently robust; this result should not be generalized to other agent auditing or process supervision methods that the paper did not test.", "section": "Safety Models Can Handle Classification, but Governance Outcomes Depend on Real-World Operation"}, {"text": "One possibility is that stable, repetitive tasks gradually adopt compact models or compilable functions, while open-ended, infrequent, or rapidly changing tasks continue to call general-purpose models. Another possibility is that new complex tasks grow faster, leaving total cloud call volumes higher even after some tasks move away. The available materials provide no data on workload composition, migration shares, or adoption rates, so they cannot establish which scenario is closer to reality.", "section": "More Local Execution Does Not Necessarily Mean Lower Cloud Demand"}, {"text": "Releasing code and checkpoints establishes only that external teams can access those assets.[S6] It does not mean that the compute, data, talent, and engineering costs required for training will fall at the same time, nor that developers will ultimately deploy the models themselves rather than buy hosted services from the original vendor or cloud providers.", "section": "More Local Execution Does Not Necessarily Mean Lower Cloud Demand"}, {"text": "If cloud services eventually offer unified deployment, scheduling, and management for different models and local modules, the scope of competition may extend from individual model calls to task execution management. But this shift depends on whether providers offer the necessary capabilities and whether customers are willing to entrust their management to a single provider; the current materials contain no corresponding product, deployment-choice, or revenue data.", "section": "More Local Execution Does Not Necessarily Mean Lower Cloud Demand"}, {"text": "Reusable local execution is therefore a technical path that can be tested further, not an established industry outcome.", "section": "More Local Execution Does Not Necessarily Mean Lower Cloud Demand"}, {"text": "On the training side, the first question is whether methods such as Muon can repeatedly achieve efficiency close to what the project team reported under compute-optimal training conditions across different model scales, hardware, and data distributions, and whether end-to-end training costs fall at the same time.[S6]", "section": "Four Types of Verifiable Signals to Watch Next"}, {"text": "On the inference side, compilable functions and remote models need to be compared on real-world task coverage, update frequency, accuracy, latency, and total maintenance costs, rather than on individual benchmarks alone.[S1]", "section": "Four Types of Verifiable Signals to Watch Next"}, {"text": "For agents and safety, attention should focus on whether independent evaluations, production adoption rates, and failure data emerge for the products listed on Zhipu's official website and the safety model described on the Qwen blog index, while distinguishing the functions shown on dynamic pages at capture time from actual operational outcomes.[S3][S5]", "section": "Four Types of Verifiable Signals to Watch Next"}, {"text": "Cloud impact must ultimately be verified through changes in the mix of calls, deployment locations, and revenue. What can currently be confirmed is that Muon has reported training-efficiency results in specific scaling experiments, and that reusable local functions, agent foundation models, and real-time safety classification have each appeared as technical or product examples; their impact on the global market and the division of business roles among the relevant vendors depends on reproducibility across settings, total costs, and the scale of adoption.[S1][S3][S5][S6]", "section": "Four Types of Verifiable Signals to Watch Next"}, {"text": "Training-efficiency replication results across models and hardware", "section": "Four Types of Verifiable Signals to Watch Next"}, {"text": "Production task coverage and maintenance costs for compilable functions", "section": "Four Types of Verifiable Signals to Watch Next"}, {"text": "Independent evaluations, adoption rates, and failure data for agents and safety models", "section": "Four Types of Verifiable Signals to Watch Next"}, {"text": "Changes in cloud and local workloads, API revenue, and deployment choices", "section": "Four Types of Verifiable Signals to Watch Next"}], "sections": [{"type": "lead", "heading": "New Developments in Training, Execution, and Safety Components; Industry Impact Depends on Reproducibility and Adoption", "paragraphs": ["The nine substantive sources used in this article include papers, technical articles, announcements, and product pages. The publication dates shown on vendor materials range from March 2025 to September 2026; all four arXiv records show updates on September 3, 2026, while Zhipu's official website and the Qwen blog index are treated as snapshots captured on September 6, 2026.[S1][S2][S3][S5][S6][S7][S8][S9][S10] On the training side, the sources cover the Muon optimizer and experiments with auxiliary representations of knowledge; on the execution side, they cover neural functions that can be saved and versioned, as well as shorter automatically optimized prompts; on the product side, they include agent foundation models, safety classification, voice interaction, and cybersecurity support for essential services. Another study tests the ability of LLM judges to identify important reasoning steps.", "These materials come from different organizations and evaluation environments. The preprints report specific benchmarks or controlled experiments, while vendor webpages provide product announcements or vendors' own descriptions of their capabilities. In particular, S3 is a captured snapshot of the Qwen blog index, and S5 is a captured snapshot of Zhipu's dynamic official website.[S3][S5] They can establish which products and technical directions the pages described at the time, but they cannot establish that the global market has broadly shifted from remote model calls to local execution, nor can the components showcased by different vendors be treated as an already integrated, unified agent product stack.", "If training-efficiency methods can be reproduced reliably and compilable functions can cover enough real-world tasks, application teams may redistribute work among general-purpose models, compact models, and local modules. The ultimate impact on cloud services will also depend on the balance between simple tasks moving away from the cloud and new complex tasks being added; these sources provide no data on workloads, the scale of adoption, or total cost of ownership."], "headers": [], "rows": [], "bullets": [], "quote": ""}, {"type": "evidence", "heading": "What the Nine Sources Establish, from Training to Interaction", "paragraphs": ["The table below treats paper-based experiments separately from product vendors' releases. Performance-related statements retain the framing used by their sources, without turning individual benchmarks, demonstrations, or vendor rankings into industry-wide conclusions.", "These materials provide no figures for global cloud market share, enterprise agent adoption, API revenue, compute procurement volumes, or production incident rates. The industry implications in the table are therefore conditional analyses, not evidence that market outcomes have already occurred."], "headers": ["Area", "Development Reported in the Source", "Limited Conclusion Supported", "Important Boundaries"], "rows": [["Training optimization", "In a technical article dated March 3, 2025, Moonshot says Muon can scale to large-model training by incorporating weight decay and adjusting the relative update scales for different parameters. Under compute-optimal training conditions, the project team reports that Muon is approximately 2 times as compute-efficient as AdamW; the project also released a distributed implementation, pretrained and instruction-tuned checkpoints, and intermediate checkpoints.[S6]", "Within the project's models, training conditions, and scaling experiments, Muon using the complete design demonstrated higher training compute efficiency than AdamW.", "The material does not separately isolate the contributions of weight decay and the adjustments to relative update scales, nor does it establish that other models, hardware, and training configurations can achieve gains of the same magnitude."], ["Knowledge learning", "An arXiv preprint updated on September 3, 2026, reports that, under a fixed token budget, redirecting some tokens from document repetition to auxiliary representations of the same knowledge can improve learning; the researchers also say the effect does not depend on the strength of the teacher model generating those representations.[S10]", "In the paper's controlled experiments, learning improved after the way knowledge was represented was changed.", "This causal result is limited to the paper's controlled conditions and cannot be used to directly infer that commercial pretraining pipelines would see the same effect; the material also provides no figures for changes in production costs."], ["Reusable neural functions", "Compile by Training uses a teacher model to generate task examples during compilation, then trains small adapters for a compact interpreter. The resulting functions do not require the teacher model at runtime and can be saved, versioned, and composed; on FuzzyBench-Hard, the paper reports 83.6% semantic accuracy and a compilation time of about one minute. The corresponding arXiv record shows an update on September 3, 2026.[S1]", "The paper demonstrates a method for converting some tasks described in natural language into local neural functions that can be run repeatedly.", "The 83.6% figure comes from a specific subset; the material provides no production task coverage figures, long-term maintenance costs, or complete cost comparison with remote interfaces."], ["Prompt optimization", "ESPO achieves an average accuracy of 74.67% across seven public natural language processing benchmarks, compared with GEPA's 70.91%; its prompts average 1004 characters, compared with 1878 characters for GEPA, making them 47% shorter. The corresponding arXiv record shows an update on September 3, 2026.[S8]", "Across these seven benchmarks, shorter prompts and higher average accuracy occurred together, but the results themselves do not prove that shortening prompts caused the accuracy improvement.", "These results do not guarantee transfer to real-world agent tasks involving tool calls, external state, and permission controls."], ["Agents and open models", "When captured on September 6, 2026, Zhipu's official website described GLM-5.2 as a flagship model that had already been open-sourced, GLM-5V-Turbo as a natively multimodal agent foundation model, and GLM-5-Turbo as a foundation model for agent scenarios with enhanced tool calling and long-horizon execution; the page also described AutoGLM as an agent model with planning, reasoning, and execution capabilities.[S5]", "At the time of capture, the dynamic page showcased open models, multimodality, and agent execution.", "This snapshot can establish only what the page displayed when captured. It cannot confirm individual release dates, prior status, continued subsequent availability, or whether the relevant rankings can be sustained over time; capabilities and rankings come from vendor descriptions and do not replace independent evaluations."], ["Real-time safety classification", "When captured on September 6, 2026, the Qwen blog index included an article dated September 23, 2025. The index text says Qwen3Guard can perform safety checks on prompts and responses, assign risk levels and categories, and cover English, Chinese, and multilingual settings.[S3]", "The article listed on the index introduces a safety guardrail model for classifying inputs and outputs.", "This source is a snapshot of a blog index, not an independently archived copy of the original article page. The page's overview of benchmark performance cannot independently verify the full evaluation results, nor can it establish a decline in incident rates or large-scale integration of the model into production systems."], ["Security support for essential services", "On September 3, 2026, OpenAI announced Daybreak for Frontline Defenders and committed $1 billion to expand access to frontier cybersecurity AI, training, and support for essential services.[S2]", "The announcement indicates that OpenAI plans to support cybersecurity applications for essential services through funding, technology, and training.", "The $1 billion figure is a commitment, not evidence that all the funds have already been spent, and it does not directly establish security outcomes, the number of participating organizations, or changes in customer relationships."], ["Voice interaction", "MiniMax's page lists January 23, 2026, as the release date for Speech 2.8 and says the model supports native tone tags and can clone a voice using a 10-second original audio clip; the page also discusses audio quality and cross-language performance.[S7]", "MiniMax positions tone control, voice cloning, and cross-language performance as product directions for this speech model.", "Naturalness, voice similarity, audio quality, and cross-language performance are presented mainly through vendor descriptions and examples; the material provides no independent quantitative evaluation."], ["Reasoning-step evaluation", "An arXiv preprint updated on September 3, 2026, defines the importance of a reasoning step as the change in expected reward after adding that step and estimates it through Monte Carlo rollouts. The experiments find that stronger LLM judges can outperform a prevalence baseline but remain well below the noise ceiling.[S9]", "In the tasks defined by the paper, readable chain-of-thought text does not mean that LLM judges can accurately identify which steps matter more to the final outcome.", "The result applies only to the paper's method for identifying step importance and cannot be used to assess all process supervision, chain-of-thought auditing, or agent auditing methods."]], "bullets": [], "quote": ""}, {"type": "mechanism", "heading": "Training Efficiency Is Influenced by Both Optimization Algorithms and Data Representation", "paragraphs": ["The Muon project focuses its design adjustments for large-scale training on two areas: incorporating weight decay and more precisely adjusting the relative update scales for different parameters. The project team says these adjustments improve training stability; in scaling experiments using the complete approach under compute-optimal training conditions, Muon was approximately 2 times as compute-efficient as AdamW.[S6] The material does not separately quantify each adjustment's contribution, so the overall efficiency result cannot be attributed to either one alone or directly extrapolated to all models, hardware, and training configurations.", "The project describes Moonlight as a 3B/16B-parameter mixture-of-experts model trained with Muon on 5.7 trillion tokens, and it released code, pretrained and instruction-tuned checkpoints, and intermediate checkpoints.[S6] These assets give external teams access to the project's implementation and models from different training stages, but actual replication still requires the corresponding compute, data processing, and engineering resources.", "The knowledge-learning experiments offer another route. Under a fixed token budget, the researchers replaced some tokens used to repeat documents with auxiliary representations of the same knowledge and observed improved learning in controlled experiments.[S10] This shows that, under the conditions defined in the paper, the way knowledge is represented affects learning outcomes, but it does not directly imply equivalent gains in commercial pretraining pipelines. The date shown for S10 is the arXiv update date, not the paper's initial publication date.", "If Muon's results can be reproduced across model scales, hardware, and data distributions, the same compute budget might support more training attempts or a larger individual training run; how that budget is allocated would still depend on a team's resources and goals. If the gains depend on a specific architecture or training configuration, the direct role of releasing code and checkpoints is primarily to make the implementation and model assets more accessible, rather than to lower training costs universally."], "headers": [], "rows": [], "bullets": [], "quote": ""}, {"type": "analysis", "heading": "Compilable Functions May Reduce Some Repeated Calls, Provided Tasks Are Stable Enough", "paragraphs": ["Compile by Training addresses a specific class of problems: tasks that are easy to describe in natural language but difficult to encode as handwritten rules, where calling a large remote model each time incurs recurring costs, latency, and vendor dependence. The method calls a teacher model during compilation to generate task examples and trains small adapters for a compact interpreter; the teacher model is no longer required at runtime, but local inference still has to be performed.[S1]", "This arrangement concentrates teacher-model calls and task-adaptation costs in the compilation stage rather than eliminating all runtime costs. In the settings reported by the paper, compilation takes about one minute, and the resulting functions can be saved, versioned, and composed.[S1] If a task is called frequently, its input boundaries are stable, and the combined costs of local inference, recompilation, and maintenance are manageable, this approach may be more suitable than making a remote call each time; infrequent or frequently changing tasks may not be suitable.", "ESPO first diagnoses error structure, then generates candidate prompts using multiple strategies, and finally performs stability selection. Across seven benchmarks, it exceeds GEPA by an average of 3.76 percentage points while using prompts that are 47% shorter.[S8] Both results occurred together in these experiments, but this does not establish that shorter prompts caused the accuracy improvement. It shows only that, on these benchmarks, achieving higher average accuracy does not require continually adding rules and exceptions.", "These experiments do not establish that remote-interface call volumes or revenue have already been affected. Some remote calls may decline only if compilable methods cover a large number of production tasks and the combined costs of local inference, deployment, updates, monitoring, and error handling are lower than those of remote calls. The sources above provide no figures for the share of tasks migrated, total cost of ownership, or changes in revenue."], "headers": [], "rows": [], "bullets": [], "quote": ""}, {"type": "analysis", "heading": "Different Vendors Showcase Different Components, but They Cannot Be Treated as a Single Product Stack", "paragraphs": ["Different vendors showcase capabilities at different stages: Moonshot released a training method, code, and checkpoints; the snapshot of Zhipu's official website showcases open-source models, tool calling, and multimodal agent foundation models; the article listed on the Qwen blog index introduces a model for input and output safety classification; and MiniMax's page introduces a speech model with tone control and voice cloning.[S3][S5][S6][S7]", "These releases and webpages show that different companies separately offer or introduce training, execution, safety, and interaction components. The materials do not show that the components already interoperate, use common standards, or have been jointly deployed, so they cannot be treated as a single product stack or used to judge how quickly the supply or market adoption of these products is changing.", "At the time of capture, Zhipu's official website also displayed open-source models, API services, model fine-tuning, AI search tools, and application-level interfaces side by side.[S5] This establishes only that the dynamic page presented open models and hosted services together at that time. It does not prove that open models will necessarily weaken its API business, nor that commercial services will become its main source of revenue. The snapshot also cannot confirm the relevant GLM models' individual release dates, prior status, continued subsequent availability, or whether their rankings can be sustained over time.", "If enterprises can migrate smoothly between models, tools, and deployment environments, open assets may expand their options. If migration remains constrained by format, data, engineering, and operational requirements, the original vendor's hosted services may still offer convenience. The materials above contain no data on migration costs, deployment choices, or revenue, so they cannot establish which of these effects is stronger."], "headers": [], "rows": [], "bullets": [], "quote": ""}, {"type": "analysis", "heading": "Safety Models Can Handle Classification, but Governance Outcomes Depend on Real-World Operation", "paragraphs": ["The article listed in the captured Qwen blog index describes Qwen3Guard as capable of classifying both prompts and responses and outputting risk levels and categories.[S3] Under this functional definition, if a system integrates the model into its generation workflow, it can be used to classify the textual safety of inputs and outputs; the index does not explain how specific systems integrate it or provide deployment-scale figures.", "Whether safety classification can become a reliable infrastructure component depends on whether it can be continuously integrated into production systems and operate with acceptable false-positive rates, false-negative rates, and latency. The available material provides no adoption rates, changes in incidents, or actual operating costs, so it can confirm only the functional direction introduced by the article listed on the index, not governance outcomes.", "OpenAI committed $1 billion to cybersecurity AI, training, and support for essential services.[S2] The commitment indicates a spending plan, but the material does not state the progress of fund disbursement, the number of participating organizations, or security outcomes, and it cannot establish a change in market power.", "When reasoning traces are used for auditing, it is also necessary to distinguish whether the text is readable from whether a judge model can identify the steps that genuinely affect the outcome. The experiments in S9 show that, in its step-importance identification task, stronger LLM judges outperform the prevalence baseline but remain well below the noise ceiling.[S9] Thus, under similar experimental settings, judging which steps are more important solely from chain-of-thought text may not be sufficiently robust; this result should not be generalized to other agent auditing or process supervision methods that the paper did not test."], "headers": [], "rows": [], "bullets": [], "quote": ""}, {"type": "counterargument", "heading": "More Local Execution Does Not Necessarily Mean Lower Cloud Demand", "paragraphs": ["One possibility is that stable, repetitive tasks gradually adopt compact models or compilable functions, while open-ended, infrequent, or rapidly changing tasks continue to call general-purpose models. Another possibility is that new complex tasks grow faster, leaving total cloud call volumes higher even after some tasks move away. The available materials provide no data on workload composition, migration shares, or adoption rates, so they cannot establish which scenario is closer to reality.", "Releasing code and checkpoints establishes only that external teams can access those assets.[S6] It does not mean that the compute, data, talent, and engineering costs required for training will fall at the same time, nor that developers will ultimately deploy the models themselves rather than buy hosted services from the original vendor or cloud providers.", "If cloud services eventually offer unified deployment, scheduling, and management for different models and local modules, the scope of competition may extend from individual model calls to task execution management. But this shift depends on whether providers offer the necessary capabilities and whether customers are willing to entrust their management to a single provider; the current materials contain no corresponding product, deployment-choice, or revenue data.", "Reusable local execution is therefore a technical path that can be tested further, not an established industry outcome."], "headers": [], "rows": [], "bullets": [], "quote": ""}, {"type": "outlook", "heading": "Four Types of Verifiable Signals to Watch Next", "paragraphs": ["On the training side, the first question is whether methods such as Muon can repeatedly achieve efficiency close to what the project team reported under compute-optimal training conditions across different model scales, hardware, and data distributions, and whether end-to-end training costs fall at the same time.[S6]", "On the inference side, compilable functions and remote models need to be compared on real-world task coverage, update frequency, accuracy, latency, and total maintenance costs, rather than on individual benchmarks alone.[S1]", "For agents and safety, attention should focus on whether independent evaluations, production adoption rates, and failure data emerge for the products listed on Zhipu's official website and the safety model described on the Qwen blog index, while distinguishing the functions shown on dynamic pages at capture time from actual operational outcomes.[S3][S5]", "Cloud impact must ultimately be verified through changes in the mix of calls, deployment locations, and revenue. What can currently be confirmed is that Muon has reported training-efficiency results in specific scaling experiments, and that reusable local functions, agent foundation models, and real-time safety classification have each appeared as technical or product examples; their impact on the global market and the division of business roles among the relevant vendors depends on reproducibility across settings, total costs, and the scale of adoption.[S1][S3][S5][S6]"], "headers": [], "rows": [], "bullets": ["Training-efficiency replication results across models and hardware", "Production task coverage and maintenance costs for compilable functions", "Independent evaluations, adoption rates, and failure data for agents and safety models", "Changes in cloud and local workloads, API revenue, and deployment choices"], "quote": ""}], "citations": [{"publisher": "arXiv", "title": "Compile by Training: Turning Natural-Language Specifications into Local Neural Functions", "url": "https://arxiv.org/abs/2609.04199v1", "published_at": "2026-09-03T17:59:49Z", "source_type": "research-paper"}, {"publisher": "OpenAI", "title": "Daybreak for Frontline Defenders: $1B to protect essential services", "url": "https://openai.com/index/daybreak-for-frontline-defenders", "published_at": "2026-09-03T13:15:00Z", "source_type": "publisher-news"}, {"publisher": "Alibaba Qwen", "title": "Blog | Qwen", "url": "https://qwenlm.github.io/blog/", "published_at": "2026-09-06", "source_type": "Official Blogs of Chinese Large Model Vendors"}, {"publisher": "DeepSeek", "title": "DeepSeek API Docs", "url": "https://api-docs.deepseek.com/news/", "published_at": "2026-09-06", "source_type": "Official Releases from Chinese Large Model Vendors"}, {"publisher": "Zhipu AI", "title": "Z.ai - Inspiring AGI to Benefit Humanity", "url": "https://www.zhipuai.cn/news", "published_at": "2026-09-06", "source_type": "Official News from Chinese Large Model Vendors"}, {"publisher": "Moonshot AI", "title": "The First Large-Scale Training Application of the Muon Optimizer - Kimi API Open Platform", "url": "https://platform.kimi.com/docs/changelog/changelog/moonlight", "published_at": "2026-09-06", "source_type": "Official Changelogs for Chinese Large Model Platforms"}, {"publisher": "MiniMax", "title": "MiniMax Speech 2.8: Bringing a \"Human Touch\" to AI Voices - MiniMax News | MiniMax", "url": "https://www.minimaxi.com/news/minimax-speech-28", "published_at": "2026-09-06", "source_type": "Official News from Chinese Large Model Vendors"}, {"publisher": "arXiv", "title": "ESPO: Error-Structured Prompt Optimization via Diagnose, Diversify, and Stabilize", "url": "https://arxiv.org/abs/2609.04197v1", "published_at": "2026-09-03T17:59:37Z", "source_type": "research-paper"}, {"publisher": "arXiv", "title": "Legibility is Not Interpretability: Comparing Judged and Actual Importance in Chain-Of-Thought Reasoning", "url": "https://arxiv.org/abs/2609.04194v1", "published_at": "2026-09-03T17:59:08Z", "source_type": "research-paper"}, {"publisher": "arXiv", "title": "Knowledge Acquisition During Pre-training? Large Language Models Learn Better With Auxiliary Views", "url": "https://arxiv.org/abs/2609.04180v1", "published_at": "2026-09-03T17:57:02Z", "source_type": "research-paper"}], "knowledge": {"question": {"en": "What do Muon’s reported training gains, reusable local neural functions, and Agent and safety-model releases establish—and under what conditions might remote-model workloads change?", "zh": "Muon报告的训练收益、可复用本地神经函数以及Agent与安全模型发布说明了什么，又在什么条件下可能改变远程模型工作负载？"}, "audience": {"en": "Model-training engineers and application architects evaluating optimizer experiments, local execution, Agent foundations, and safety classification.", "zh": "评估优化器实验、本地执行、Agent基座与安全分类的模型训练工程师和应用架构师。"}, "decision": {"en": "Identify which results merit replication or production evaluation, without treating benchmark gains and vendor announcements as evidence of lower total costs or declining cloud demand.", "zh": "判断哪些结果值得复现或开展生产评估，而不把基准收益和厂商公告当作总成本下降或云端需求减少的证据。"}, "contribution": {"en": "The article separates experimental results from vendor descriptions and conditional industry analysis. It explains Muon’s compute-efficiency comparison, Compile by Training’s shift of teacher-model work to compilation, and the distinct Agent and safety components presented by Zhipu AI and Qwen.", "zh": "文章区分实验结果、厂商描述和有条件的产业分析，解释Muon的计算效率比较、Compile by Training将教师模型工作集中到编译阶段的安排，以及智谱和Qwen分别展示的Agent与安全组件。"}, "limitation": {"en": "The evidence comes from specific experiments, preprints, announcements, and webpage snapshots. It does not establish cross-setting reproducibility, production adoption, total ownership costs, safety outcomes, or changes in cloud workloads and revenue; the separately presented components are not a demonstrated integrated product stack.", "zh": "证据来自特定实验、预印本、公告和网页快照，未能证明跨场景复现、生产采用、总拥有成本、安全效果或云端工作负载与收入变化；分别展示的组件也不构成已经验证的统一产品栈。"}, "nodes": [{"key": "n1", "kind": "entity", "name": {"en": "Muon", "zh": "Muon"}, "mention": "Muon", "id": "entity-507cbee575ba94145df9"}, {"key": "n2", "kind": "entity", "name": {"en": "AdamW", "zh": "AdamW"}, "mention": "AdamW", "id": "entity-99a90f03463afe01bf82"}, {"key": "n3", "kind": "entity", "name": {"en": "Compile by Training", "zh": "Compile by Training"}, "mention": "Compile by Training", "id": "entity-90607329e06697946a72"}, {"key": "n4", "kind": "concept", "name": {"en": "Remote-model calls", "zh": "远程模型调用"}, "mention": "远程模型调用", "id": "concept-2e6b6c987c255d8fb561"}, {"key": "n5", "kind": "entity", "name": {"en": "Qwen3Guard", "zh": "Qwen3Guard"}, "mention": "Qwen3Guard", "id": "entity-14379eba60976628b0b8"}, {"key": "n6", "kind": "concept", "name": {"en": "Safety classification of prompts and responses", "zh": "提示和回答的安全分类"}, "mention": "输入和输出的文本安全分类", "id": "concept-90aed27dba4fd4b66ed3"}], "edges": [{"source": "n1", "target": "n2", "relation": "compares", "stance": "reported", "explanation": {"en": "Moonshot AI reports approximately twice AdamW’s compute efficiency for the complete Muon design under compute-optimal training conditions. The result neither isolates the contributions of individual modifications nor establishes the same gains across other models, hardware, and configurations.", "zh": "Moonshot AI报告，在计算量最优训练条件下，采用整体设计的Muon计算效率约为AdamW的两倍。该结果既未隔离各项调整的贡献，也未证明其他模型、硬件和配置能获得同等收益。"}, "quote": {"en": "The project team states that these adjustments improved training stability; in scaling experiments using the complete approach under compute-optimal training conditions, Muon’s compute efficiency was approximately twice that of AdamW.", "zh": "项目方称，这些调整提高了训练稳定性；在采用整体方案且处于计算量最优训练条件的扩展实验中，Muon的计算效率约为AdamW的2倍。"}, "citations": [6], "sourceId": "entity-507cbee575ba94145df9", "targetId": "entity-99a90f03463afe01bf82"}, {"source": "n3", "target": "n4", "relation": "changes", "stance": "hypothesis", "explanation": {"en": "The author conditionally argues that Compile by Training could reduce some remote calls by replacing repeated teacher-model use with reusable local inference. This requires broad production-task coverage and lower total costs, including deployment, updates, monitoring, and error handling; no reduction in calls or revenue is demonstrated.", "zh": "作者有条件地推测，Compile by Training通过可复用本地推理替代逐次使用教师模型，可能减少部分远程调用。前提是覆盖大量生产任务，且包含部署、更新、监控和错误处理的总成本更低；文章未证明调用量或收入已经下降。"}, "quote": {"en": "These experiments have not demonstrated that remote-interface call volume or revenue has already been affected. Only when compilable methods cover a large number of production tasks, and the total cost of local inference, deployment, updates, monitoring, and error handling is lower than that of remote calls, might some remote calls decrease.", "zh": "这些实验没有证明远程接口的调用量或收入已经受到影响。只有当可编译方法覆盖大量生产任务，而且本地推理、部署、更新、监控和错误处理的总成本低于远程调用时，部分远程调用才可能减少。"}, "citations": [1], "sourceId": "entity-90607329e06697946a72", "targetId": "concept-2e6b6c987c255d8fb561"}, {"source": "n5", "target": "n6", "relation": "enables", "stance": "reported", "explanation": {"en": "The captured Qwen blog listing describes Qwen3Guard as detecting safety issues in prompts and responses and assigning risk levels and categories across English, Chinese, and multilingual settings. This is a reported product capability, not independent evidence of production adoption or fewer incidents.", "zh": "Qwen博客汇总页快照将Qwen3Guard描述为可对提示和回答进行安全检测，并在英语、中文及多语言环境下给出风险等级与类别。这是材料所载的产品能力描述，并非生产采用或事故减少的独立证据。"}, "quote": {"en": "The blog listing states that Qwen3Guard can perform safety detection on prompts and responses, provide risk levels and categories, and cover English, Chinese, and multilingual settings.", "zh": "汇总页文字称，Qwen3Guard可对提示和回答进行安全检测，给出风险等级和类别，并覆盖英语、中文及多语言环境。"}, "citations": [3], "sourceId": "entity-14379eba60976628b0b8", "targetId": "concept-90aed27dba4fd4b66ed3"}]}, "knowledgeMap": "/knowledge", "license": {"model": "open", "price": 0.002, "currency": "USDC", "paymentProtocol": "x402", "network": "eip155:84532", "payTo": "0x23a885262A89c32eB582f4439184769dA22c467c", "variants": {"A": "/agent/v1/articles/ai-research-20260906-0025-766?lang=en", "B": "/agent/v1/articles/ai-research-20260906-0025-766/paid?lang=en"}}, "accessVariant": "A", "contentPolicy": {"citationAllowed": true, "attributionRequired": true, "trainingUse": "contact publisher"}}