This revision retains the original article's focus on model projects, product positioning, openness, and evidentiary limits. It replaces the menu excerpt with the detailed Baichuan interface documentation saved on September 13 and incorporates Baidu's three Agent categories and the conditions for GPU-CFR's static computation. Linking tool requests to results can help analyze execution coordination, but does not demonstrate task success. Cost and migration serve only as supplementary dimensions of comparison, not as grounds for identifying market winners.
Revision Note: Different Materials Answer Different Questions
Revised on September 14, 2026, with information current through September 13; the initial publication date of the original article 197 remains unchanged. Materials marked as published on September 14 are listed only as subsequent supplements and are excluded from factual assessments as of September 13. This revision uses the saved detailed Baichuan interface documentation to correct the earlier selection of a menu excerpt and incorporates evidence-supported analysis of mechanisms from the earlier article 177.
These materials cover model titles, vendor product pages, interface specifications, research abstracts, and industry projects; they cannot be combined into a single capability ranking. Meta AI presents application directions for its models, Hugging Face publishes articles on encoders and workflows, Baidu presents the division of roles among its Agent products, and Baichuan specifies message and tool interfaces. They establish public positioning, but do not directly demonstrate adoption scale, reliability, or productivity.[S1][S2][S6][S7][S8][S9][S10][S11]
Model and Project Positioning: Names Cannot Be Converted into Rankings
The two Meta AI records dated September 8 contain only a “FEATURED” label, a date, and a link, with no content summary. Product functions, modes of access, or scope of availability cannot be inferred from the URLs.[S1][S10] The titles of two projects from the same day connect Meta models to assistive robotics at the University of Pittsburgh and to Genesis Mission projects, respectively. These establish application directions, but not safety, research efficiency, or deployment scale.[S8][S9]
Hugging Face published a NeoMME blog post on September 3 whose title positions it as an efficient, multimodal-native, multilingual encoder.[S2] “Efficient” is the project's own characterization. A specialized encoder approach could offer deployment value only if comparisons at equivalent quality and on equivalent hardware show lower costs. This is a conditional analysis, not a conclusion about relative efficiency established by the available materials.
Baidu's page covers roles including all-modality processing, vision-language processing, document parsing, speech, long context, code reasoning, and Agents.[S6] The vendor states that ERNIE 5.1's pretraining cost is “6% of the industry level” and that GLM-5.3 delivers a 50% improvement in the programming experience over its predecessor. However, the supplied text does not provide sufficiently consistent baselines, tasks, hardware, or evaluation methods to cite these directly as comparative advantages. A large number of model listings does not establish capability leadership either.
The Baidu entry is dated July 22, but the collected text mixes promotional activities with product navigation, so the entire page cannot be interpreted as a single comprehensive update on that date. Baichuan's September 13 record likewise provides no version history; the date on which the material was saved must be distinguished from the launch date of each capability.[S6][S7]
Baidu's Three Agent Categories: Planning, Orchestration, and Collaboration Are Not Synonymous
Baidu describes Qianfan as “a one-stop enterprise-grade large-model service platform centered on Agents” and explicitly distinguishes three development approaches: autonomous-planning Agents think and plan independently for basic or routine business operations; workflow Agents conduct conversations strictly according to human-orchestrated processes for business operations requiring rigor; and multi-agent collaboration supports importing custom sub-Agents and using multiple officially preconfigured sub-Agents to handle complex scenarios collaboratively.[S6] This is the vendor's product classification, not a reliability ranking.
In terms of mechanisms, autonomous planning delegates some step selection to the system at runtime; workflows encode more of the sequence and branching in advance; and multi-agent collaboration adds subtask allocation and the consolidation of results. These are interpretations based on the page's positioning, not a basis for inferring its actual scheduling implementation. Complex tasks do not necessarily require multiple Agents, and predefined workflows may also call multiple tools.
The page also presents MCP components, search, encyclopedia, PPT tools, featured Agents, application deployment, and an AI Workflow entry point.[S6] Listing these side by side describes how offerings are organized, but does not demonstrate deep interoperability among components. Agents may be able to coordinate execution only if model generation, context provision, tool connections, and process constraints work together. Unified navigation may instead merely offer purchasing convenience or brand integration; it does not prove growth in computing demand or that customers are already locked in.
Baichuan's Detailed Protocol: Linking Tool Messages Does Not Mean Completing External Operations
The Baichuan material saved as S7 contains detailed interface documentation, not just a menu. The documentation specifies POST requests to https://api.baichuan-ai.com/v1/chat/completions, using JSON requests and Bearer APIKey authentication, with model and chronologically ordered messages submitted. Roles include user, assistant, system, and tool; a system message must be the first entry in the conversation history. These are request-format conventions, not records of service availability.[S7]
The key distinction for tool messages is that tool_calls represents tool-call messages generated by the model and is permitted only for the assistant role; under that role, content and tool_calls are mutually exclusive. tool_call_id is used in records of tool-call results and is populated only for the tool role.[S7] Ordinary answers, tool requests, and returned results therefore cannot be treated as the same kind of message. Nor does the table's designation of content as required justify ignoring its explicit mutual-exclusion condition with tool_calls.
At the mechanism level, an application can use these conventions to identify when the model is requesting a tool and use call identifiers to organize returned results and continue the conversation. However, the supplied excerpt does not fully show the rules for generating call identifiers, concurrent returns, failure retries, or handling duplicate operations, so these cannot be filled in as confirmed protocol details. Parseable requests and linkable results do not demonstrate correct function execution, successful changes to external state, or completion of the overall task.
tools lists three categories—retrieval, web_search, and function—and supports multiple function entries at the same time. The documentation recommends not using function together with search or retrieval. When search and a knowledge base are both configured, the knowledge base takes priority and search does not take effect. Once the knowledge base is enabled, only questions are supported as input, not complex instructions such as scenario settings. In knowledge-base-only mode, no answer is given if nothing is retrieved, whereas the default allows the model to generate its own output. Web search incurs an additional charge.[S7] These restrictions affect context sources, answer boundaries, and costs, and cannot be glossed over by saying that the system “supports tools.”
function provides a name, a description, and parameter definitions; the description helps the model choose when and how to call it. The parameter field is labeled string, but the explanation calls it a JSON-formatted object. The excerpt is insufficient to resolve this discrepancy. Feature support must also be checked individually: the function support list includes Baichuan2-Turbo but not Baichuan3-Turbo-128k, while the JSON mode list does the reverse.[S7] The overall model list cannot be treated as the support list for every feature.
JSON mode also requires a system or user message requesting JSON generation, and output may be truncated when finish_reason is length.[S7] These are usage conditions described in the documentation, not instructions being executed by this article. Even if output conforms to JSON syntax, that does not establish that the parameters are semantically valid or that tool execution succeeded.
Openness and Industry Outcomes Must Be Verified Separately
Public descriptions, applications for API access, downloadable weights, licenses permitting modification and redistribution, and third-party reproducibility represent different levels of openness. The M3Plus API application entry point on Baichuan's page does not demonstrate that weights are open; Meta's project titles do not establish licensing terms; and NeoMME's publication on Hugging Face is no substitute for model files and license text.[S1][S2][S7][S8][S9][S10]
Hugging Face's September 10 article is titled “Rebuilding AUTOMATIC1111 with Gradio Workflow.” The supplied materials contain only the title and metadata, which are insufficient to establish implementation, compatibility, adoption, or migration outcomes.[S11] Model replaceability and workflow portability must still be distinguished: similar interfaces do not guarantee that configurations, business rules, and tool connections can be transferred directly.
Industry projects likewise cannot substitute goals for results. MIT News states that people affiliated with MIT are collaborating with the MIT-IBM Computing Research Lab to bring rigorous theory into production systems. OpenAI, AIRPPU, and WAN-IFRA announced a program for Ukrainian news organizations, with goals involving innovation, resilience, and independent journalism.[S3][S5] Neither establishes that production deployment or journalism quality has already improved.
The Diffusion TV abstract describes using the antenna of a modified CRT television to adjust the clarity of generated images and sound, and knobs to switch among animal channels from the past, present, and future. The authors position it as an embodied experience of explainable AI; this is not a measurement of improved user understanding.[S4] AI-GUIDE, covered in a subsequent MIT report, was jointly developed by Lincoln Laboratory and Massachusetts General Hospital; improving casualty health remains a vision stated in the summary.[S12] These projects also remind readers that AI is delivered through more than cloud-based chat.
GPU-CFR: Gains from Fixed Computational Structures Cannot Be Extrapolated to Dynamic Agents
The GPU-CFR abstract discusses counterfactual regret minimization: for a fixed game, the structure of iterations can be determined in advance, with only numerical values changing. The authors compile it into static dataflow, precompute indices, and batch operations by depth. Because shapes, indices, and buffer addresses remain unchanged, one iteration can be recorded and then replayed using CUDA Graph Replay.[S13]
The authors report up to an 18.1-fold reduction in the number of framework operations; this is not an 18.1-fold improvement in general task speed or reduction in cost. The title's “80x Faster” likewise cannot be generalized beyond its comparison conditions. The supplied abstract is truncated at “On one A100,” leaving insufficient information to reconstruct the full baseline.[S13] The only mechanism-level takeaway is that reducing repeated scheduling may be effective when the structure is genuinely fixed. Dynamic Agents change paths and tools based on intermediate results; repetition in a business process does not mean that underlying shapes and addresses remain fixed.
Cost and Migration Are Supplementary Dimensions, Not Substitutes for Model Capability
If a model fails to meet the quality threshold, integration cannot substitute for capability improvements. Only when quality, safety, and service scope are equivalent is it appropriate to compare cost per successful task, accounting for calls, additional searches, integration and maintenance, failure recovery, and human review, alongside task completion rates, latency, and human takeover. Comparing only the price of a single answer is insufficient.
Migration should be assessed by the actual work required to export configurations, rewrite rules, reconnect tools, and perform necessary regression testing. Convenience can translate into lock-in only if a platform both reduces ongoing maintenance and creates dependencies that are difficult to replace; replaceable components may interrupt this mechanism. This article conducted no tests and found no evidence of profit growth or customer lock-in. The limited sample cannot represent the direction of the global market or companies' competitive positions, and the discussion does not constitute investment advice.
Key Source Links and Scope of Verification
The links below correspond to the existing source numbers. The strength of each citation is limited to the titles, abstracts, or text excerpts supplied for this revision; the existence of a link is not a substitute for substantive evidence.
| Source | URL | Scope of Material |
|---|---|---|
| [S2] Hugging Face | https://huggingface.co/blog/Hcompany/neomme | NeoMME title and metadata |
| [S3] MIT News | https://news.mit.edu/2026/from-mit-to-ibm-expediting-ai-and-quantum-deployment-0902 | Summary of MIT's collaboration with the IBM lab |
| [S4] arXiv | https://export.arxiv.org/abs/2609.05404v1 | Diffusion TV design and the authors' positioning |
| [S5] OpenAI | https://openai.com/index/supporting-independent-journalism-in-ukraine | Summary of the announcement of the Ukrainian journalism program |
| [S6] Baidu AI Cloud | https://cloud.baidu.com/product/qianfan_home/token-plan-activity.html | Text covering product positioning, three Agent categories, and navigation |
| [S7] Baichuan AI | https://platform.baichuan-ai.com/docs/api | Detailed interface documentation saved on September 13 |
| [S8] Meta AI | https://ai.meta.com/blog/assistive-robotics-university-of-pittsburgh-sam-dino/ | Assistive robotics project title |
| [S9] Meta AI | https://ai.meta.com/blog/genesis-mission-lawrence-berkeley-national-laboratory-segment-anything-dino/ | Genesis Mission project title |
| [S11] Hugging Face | https://huggingface.co/blog/gradio-workflow-1111 | Workflow article title and date |
| [S12] MIT News | https://news.mit.edu/2026/lifesaving-lincoln-laboratory-technology-wins-tech-transfer-award-0911 | Summary identifying AI-GUIDE's developers and its vision for health |
| [S13] arXiv | https://export.arxiv.org/abs/2609.11923v1 | GPU-CFR's static computation conditions and partial quantitative results |
Subsequent Supplement: Excluded from Assessments as of September 13
S14 is marked as published on September 14, but was collected on September 13. This chronological inconsistency means it cannot be regarded as having been formally published within the observation window. Its summary states that Perplexity uses Astra to handle communications, software modifications, and production monitoring, with less frequent checks than earlier models required. This is recorded here only as OpenAI's account of the case, not as evidence of task reliability, cost improvements, or successful autonomous execution.
| Source | URL |
|---|---|
| [S14] OpenAI | https://openai.com/index/perplexity-improving-accuracy-with-astra |
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