Following the release of the decision model Jev, overseas tech giants have quickly followed suit. OpenAI released the Decisions API within two weeks, closely followed by AWS launching Strands Decider 2B, and subsequently Microsoft releasing Microsoft-Decision-1. Among domestic model giants, only Alibaba Cloud has listed decision-model-preview, which is currently in a limited-time free phase.
Looking solely at this comparison, the conclusion is that domestic models are one step behind again. However, Ai Fenxi believes that the temporary decision of domestic model giants not to follow suit is a rational one.
01. What Innovations Has Jev Introduced?
First, let us look at what the decision model represented by Jev is. For example, when a customer complaint ticket comes in, three things need to be judged: Which department should the ticket be routed to? How urgent is the ticket? Does the customer need a refund? The model specifically designed to make these three micro-judgments is the decision model. These three judgments share an input context and only need to output clear judgment results, without requiring a complete reasoning process or natural language responses.
Due to the output characteristics of these judgments, using current Large Language Models (LLMs) to answer presents the following three issues.
First, the output speed is slow. When generating answers, an LLM must output tokens one by one, with each token depending on the previous one, which leads to a slow output speed.
Second, the decision-making cost is high. Each output from an LLM generates a text with explanations, but the actual decision results do not need these explanatory texts, resulting in a lot of redundant costs during the decision-making process. In particular, generating each token by an LLM requires re-invoking the context, so the output price is generally much more expensive than the input price, which further drives up the decision-making cost.
Third, the decision results are unreliable. During the conversion from the text output by the LLM to the judgment results required for decision-making, structural violations may occur, meaning the format deviates. According to public tests, under unconstrained conditions, the probability of violations for large-parameter models is about 6%, and for small-parameter models, it is 18%. During the execution of long-horizon tasks by Agents, such micro-judgments may need to be made hundreds or even thousands of times.
The above three issues, when amplified hundreds or thousands of times, will cause significant deviations in speed, cost, and results. Therefore, a specific category of models was born to solve these problems. Taking Jev as an example, the input-side model receives a context and a set of questions, each with several fixed options. The output-side model only requires a single computation, returning the probability distribution and confidence level falling on the options for each question, without returning any text.
Figure 1: Jev Model
In process field tests, Jev has significantly improved performance. According to official statements, its speed is 193.6x faster than that of LLMs, and its cost is 444.6x lower, while structurally eliminating the issue of structural violations at the architectural level. The industry defines this type of model as a Typed Decision Model (hereinafter referred to as TDM). TDM is not a new model category created by Jev; its history can be traced back to the early BERT classifiers.
Based on exchanges between Ai Fenxi and industry experts, Jev's technological innovation is not overly complex, but through engineering, it has achieved the extreme in terms of speed and cost. In the narrative of overseas model giants pursuing AGI and capability breakthroughs, Jev's approach of competing on the cost curve rather than the capability curve is considered an outlier.
02. Overseas Vendors Follow Up Quickly
Due to the low technical complexity and low follow-up costs, overseas vendors have also acted quickly. Microsoft, AWS, and Cloudflare, mentioned at the beginning, all conducted post-training based on small-sized Qwen models, launching their own TDMs successively within one month. Some smaller overseas vendors have also constructed new TDMs based on the previous generation of BERT models.
Figure 2: Overview of Overseas Vendors' TDMs
OpenAI, on the other hand, did not do any post-training at all. It directly added the Decisions interface to the small model Luna, constrained the output, and made specific optimizations for the speed of result return. The response speed officially disclosed by OpenAI is 150 milliseconds, which is close to Jev's official 114 milliseconds.
In terms of pricing, OpenAI's pricing is USD 0.1 per million tokens for input, with output being free of charge. Although slightly higher than Jev's USD 0.042, OpenAI additionally supports image input, offering more modalities than Jev's pure text input. The motivations for overseas vendors to follow up on TDM can be categorized into two types.
Figure 3: Motivations of Vendors Following Up on TDM
The first motivation is that there are a large number of internal decision-making scenarios requiring the use of TDM, with substantial demand and practice in scenarios such as advertising and risk control. For example, Cloudflare, a company originating from CDN, has a massive number of decision-making scenarios at the edge, including scheduling, security, and content moderation, making TDM a rigid demand for internal use.
In its past business, Cloudflare has also accumulated a large amount of historical decision-making data. Therefore, during the post-training process, Cloudflare first used its own accumulated real data, followed by synthetic data. In contrast, Jev does not have an accumulation of decision-making data from historical scenarios and entirely used synthetic data for post-training.
The second motivation is to sell model APIs externally, where the richer the product shelf, the better. To this end, OpenAI and others supplemented the TDM category on their shelves at the very first opportunity.
03. Follow-up Status of Domestic Models
Among domestic model giants, the only one that has entered the fray is Alibaba. Tencent, ByteDance, and others have not followed up on TDM based on the motivation of internal scenario usage or enriching the model shelf. Other followers are not top-tier vendors.
Figure 4: Overview of Domestic Vendors' TDMs
Ai Fenxi believes that the choice of domestic model giants to temporarily not follow up is not a misjudgment, but a correct decision.
First, TDM cannot support an independent category narrative. Based on the previous discussion, it can be seen that the technical architecture complexity of TDM is not high. Therefore, the level at which TDM is situated is similar to models like embedding and reranker, rather than being on the same level as world models and AI4S models.
Second, TDM does not reflect the upper limit of a model's capabilities and is unrelated to the AGI narrative. The mainstream narrative of models is still pursuing breakthroughs in capability upper limits, rather than speed and cost in single-point functions. TDM has not changed in terms of capability breakthroughs. Independent reproduction shows that the accuracy of TDM and frontier models on the same task is 67.8% vs 67.9%, a difference that is negligible.
Third, the existing decision models of domestic giants are sufficient, and replacing them yields no incremental benefits. In the internal scenarios of giants like Tencent and ByteDance, there are numerous scenarios such as advertising, risk control, and gateway routing. They have already constructed highly performing decision-making systems, meaning not just one TDM, but multiple TDMs connected in series, ensuring the decision misjudgment rate is at 1‰ or even lower. Therefore, these giants lack strong motivation to replace them. Based on the above reasons, it is easy to understand why domestic models do not have a strong willingness to follow up on TDM.
However, if TDM is just a very small model category, why has TypeSafe, the company behind Jev, reached a post-money valuation of USD 7.5 billion in its Series A? According to public statements by founder Diogo Almeida, TypeSafe plans to build a model family for machines, software, and Agents, allowing software to directly consume model outputs instead of humans reading them. In other words, TypeSafe's vision is to build a set of machine-native models parallel to OpenAI and Anthropic.
Jev changes the model output from natural language to typed decisions, but Jev is only the first link in the model family, with many more layers of models to be built above it. Almeida does not even accept the classification of decision models or TDM. He states that in the future, there will be types within machine-native models that do not solve decision-making problems.
What TypeSafe really aims to do is much bigger than TDM. TypeSafe's official statement is to make intelligence composable, allowing software to directly consume model outputs instead of humans reading them. The title of the company's manifesto is "Build Products, Not Gods."
When asked whether they plan to build a frontier lab, Almeida's answer was that the main products of frontier labs are fear and hype, and he hopes TypeSafe's product is intelligence. From this perspective, model giants do not know what comes after Jev. In a track where the endpoint has not yet been described, followers do not know what position to follow up to, making it difficult to place their bets. Therefore, model giants are not unwilling to follow up; they just cannot see clearly yet and have not taken action.