We tested Jev using 190 financial report tasks: When AI determines that it was sold as parts...

Authors | Lin Ke, Zheng Hao

In the first half of 2026, Kweichow Moutai's operating revenue was 90.7 billion yuan, an increase of less than two percentage points; net profit attributable to the parent company was 44.5 billion yuan, a decrease of two percentage points; and operating cash flow was 70.7 billion yuan, an increase of nearly five and a half times.

If you were to put this passage into an exam and be given five descriptions to choose the one that best fits the question, what would you choose?

On September 15, the American startup TypeSafe AI ended its two-year hiatus by releasing a model called Jev, igniting the attention of Silicon Valley and the AI community.

After getting the hang of it, we fed it the question above, and it selected the same description three times in a row, each time with full confidence. The whole question cost less than $0.001.

What makes this model special is that it doesn't write a single word, doesn't explain why, doesn't write code or chat. Give it a piece of material and several alternative descriptions, and it will only check one of them and then report a number to indicate how confident it is.

This simple, minimalist tool was integrated into multiple developer platforms and model gateways, including OpenRouter, within a week. Within days, an open-source clone was created, causing a stir in Silicon Valley and the AI community.

Jev founder Diogo Almeida spent four years researching at OpenAI, and his work on InstructGPT was the precursor to ChatGPT. DCVC led a $40 million seed round. The model is named after the 19th-century economist Jevons, a name that encapsulates the company's entire stake.

Returning to the question about Moutai at the beginning, we can see that the correct options almost completely reiterate the facts in the material. The question itself is not very difficult, but what is really worth asking is: how can an AI that doesn't even write a single word generate such a lot of buzz?

1. From consultant to switch

Anyone who has worked on AI products knows a common dilemma.

A company's system makes a lot of trivial judgments every day: whether this email counts as a complaint, whether this announcement mentions litigation, and whether this refund should be automatically approved.

The most common approach is to ask a large model to answer each time. However, the core working method of a large model is generation, which requires writing word by word. If you ask it a "yes or no" question, it may first write three paragraphs of analysis before giving a conclusion. It charges by word count, and you have to wait for several seconds or even tens of seconds. The returned text still needs to be broken down and verified by the program. If the format is incorrect, it has to be redone.

What's more troublesome is that when the large model says "I'm very sure," the statement itself doesn't come with a scale.

Sometimes, being "very certain" is like using "a little" in traditional cooking. Whether it's 70% or 90%, or the 1-3% that corresponds to pre-cooked ingredients, it's uncertain which direction it falls into, and there's no way of knowing.

As a result, many judgments that should have been made automatically by machines are stuck in a dilemma: using large models is too expensive and slow, while relying on hard-coded rules is not flexible enough.

Jev did something that seemed simple: he pre-wrote the answers so the model would only be responsible for selecting them.

You simply feed it the materials and answer choices, and it simultaneously assigns a support level to each option, returning the one with the highest support. Because it doesn't generate the material word by word, the speed is orders of magnitude faster; and because it only charges based on the number of words read in the material, with no output fee, the cost of each judgment is almost negligible.

The most crucial business difference here is the change in the unit of measurement.

The big model sells "word count", while Jev sells "one-time judgment".

Word count is the pricing method for services, while a single judgment is the pricing method for components. Components can be included in a procurement list, their unit economics can be calculated, and they can be crammed into a single line of code and run tens of thousands of times. If the component is cheap enough, the software will generate a large number of judgment points that were previously not worth calling AI for.

TypeSafe conducted its own evaluation, comparing Jev with several large model workflows in five business scenarios.

In terms of accuracy, Jev only caught up with the mid-range level, falling 6 percentage points behind the best. Its entire advantage lies in price and speed: each judgment is hundreds of times cheaper and dozens of times faster.

Another frequently mentioned selling point is that they "don't make things up."

Strictly speaking, Jev's answer can only fall within the options you provide, so it cannot create a concept or number out of thin air.

However, it can confidently choose the wrong option, and the form of the error has changed from "writing a non-existent sentence" to "confidently ticking the wrong option," the latter being less likely to be detected.

The founder himself admitted that it is possible to answer incorrectly when one is highly confident.

TypeSafe also did something unusual for an AI company—it listed all the known weaknesses of its model in the instruction manual. It stated that Jev reads dates like ordinary text, is not good at arithmetic, gets distracted by irrelevant content in the material, performs best in English, and is somewhat weak in Chinese. It also suggested testing with its own material in non-English scenarios first.

This instruction manual is worth writing because it precisely defines what we are going to do next.

II. Faced with the pig cycle, it hesitated.

TypeSafe's instruction manual states that the Chinese language is weaker than the English language and recommends that users test it themselves first.

Since Jev's release, there have been very few publicly available Chinese-language test results. Therefore, after resolving the API issue, we conducted a round of testing using domestic financial scenario materials.

The rules were very simple: the materials were taken from regulatory authorities and the original interim reports of various companies, and cross-checked using Wind. The reference answer for each question was written and archived before the model was called, and not a single word was changed after it was written.

Ask each question three times to see if it will change its mind on its own.

We prepared 15 questions covering a wide range of topics, from the "two new" policies to reserve requirement ratio and interest rate cuts, from coal supply shocks to automotive chip shortages, from photovoltaics to the hog cycle, plus the interim reports of nine companies: CATL, BYD, Vanke, Tencent, Moutai, Meituan, JD.com, Pop Mart, and Anta.

45 judgments, all correct, with 3 identical answers for each question.

The least interesting part of this report card is the perfect score ; the interesting part is question 15.

In August 2021, the number of live pigs in stock increased by nearly 30% year-on-year. Farmers lost an average of thousands of yuan per pig. The number of breeding sows in stock was still rising year-on-year, although it had declined month-on-month. The oversupply was not over yet, and signs of capacity contraction had appeared. This issue is right at a turning point.

Jev chose "from surplus to rebalancing" in all three rounds, but this was the only question out of 15 that he wasn't fully confident about, giving 85%, 88%, and 85% respectively in the three rounds. He left the remaining uncertainty with "still in surplus," without offering any other incorrect directions.

A full count of judgments proves nothing; this hesitation indicates that a certain degree of judgment is still being formed.

The reason for the perfect score is related to the test questions.

For example, the growth rates, such as "year-on-year growth of 54.80%" and "year-on-year decrease of 7.13%", were all calculated in advance and given to it. It never did any addition or subtraction from beginning to end.

In TypeSafe's own evaluation, Jev's worst performance was invoice verification, which required comparing amounts, quantities, and delivery dates. It only scored 61.8% of the tests. TypeSafe's manual also states: "Leave the arithmetic to the code."

We essentially created a test paper following the instructions, avoiding all its known weaknesses.

Whether a model performs well or not depends more on the design of the questions that are presented to it. This also means that the core competency of people who develop products around it may lie in the design of options and data preprocessing.

Then we did a second round of testing, giving the accounting work back to it.

3. Making the same division problem wrong 10 times in a row.

In the second round, we selected 3 companies and asked 26 questions, each of which was asked 5 times. The questions were the same as we predicted, and all of them were archived before being used.

We bet on 13 things that would fail, and in the end we only won on two.

For example, it actually counted correctly. In Pop Mart's semi-annual report, it counted several of the 11 IPs and business lines whose revenue declined. The ranges it provided in all five rounds were correct, but the longer the list, the less certain it became. TypeSafe's manual states that it cannot reliably count, but it did this.

Tsingtao Brewery's second-quarter profit declined while its cumulative profit for the first half of the year increased. These opposing figures were mixed together in the same document, and the company was not misled by the single-quarter figures five times. As for Haier's revenue growth in 2024, how much came from the three acquisitions completed at the end of the year and how much came from its own revenue? The proportion needs to be calculated by subtracting the contribution of acquisitions from the total increase, and the company got it right five times.

They also answered all the questions about the revenue growth rates of the four companies correctly. For example, regarding Yili's performance, the actual growth rate was 4.13%, and the dividing line was set at 4%, which was only 0.13 percentage points away. They were 97% confident about this question.

Submitting the entire chapter of Moutai's management discussion directly resulted in accurate answers; the English version of Pop Mart, consisting of 44,000 characters, was also submitted and answered correctly 10 times. However, the Chinese version of Haier, which was also an entire chapter of interim report text, was rejected by the interface 5 times.

It is evident that for the same length of text, Chinese text consumes significantly more billing units and is more likely to hit the reading limit.

Then the test reached the boundary of the model.

We provided CATL, Kingsoft Office, and Moutai with their revenue and costs for two years respectively, and asked whether their gross profit margins had improved or worsened.

This kind of question requires calculating the gross profit margin for two years and then comparing the changes. Even a person would need a calculator just to look at it. It's basically impossible to answer this kind of question correctly, but the questioner knows that the probability of answering it is less than 20%, with the probability almost evenly distributed among the 6 options. It's like openly admitting that they are "guessing".

It got most of the questions right if the result was far from the dividing line and the direction could be determined at a glance; however, it got two out of three questions wrong if the result was close to the dividing line and the result had to be calculated to the decimal point.

This confirms the definition in the manual: Jev is not a calculator; arithmetic must remain in the code. For those who want to use it to create financial products, this line is clearly drawn: from data retrieval and alignment to ratio calculation, these tasks must be completed by the program. Jev can only handle the final semantic judgment step.

However, an unexpected discovery during the test was that it answered a question incorrectly without realizing it was wrong.

To test its performance on the benchmark, we chose Midea's 2025 interim report for comparison: operating costs increased by 16.83% year-on-year, and our dividing line is set at 17%.

The materials listed two figures for operating costs in the same period of 2024: approximately 158.4 billion yuan (original figure) and approximately 159.9 billion yuan (adjusted figure). It selected "reaching 17%" five times, and each time with over 80% certainty.

We initially thought that the two sets of figures had misled it. If calculated according to the original figures, the growth rate would indeed reach 17.9%, just above the threshold. So we deleted the original figures and the adjusted amount, leaving only the two clean figures after the adjustment, and asked five more times.

Surprisingly, they were all wrong again, which actually increased their confidence level.

We got the same question wrong 10 times in 10 different ways. This is the third time we've guessed wrong about where we went wrong. Yili got the answer right while sticking to the line, but Midea got it wrong while sticking to the line. The reason is uncertain, mainly because Jev didn't show the solution process, making it impossible to trace the source. All we can say is that we observed the differences in the number of digits and units. For example, Midea's number has 9 digits and the unit is thousands, while Yili's has 4 decimal places and the unit is hundreds of millions.

An error in the gross profit margin question will trigger an alarm. It knows it's guessing, so users can set a threshold and switch to human intervention if the user isn't confident enough.

What we should be wary of is the issue of beauty itself, which is very certain that it is right. In the context of real products, the latter is more dangerous.

What's particularly intriguing is what exactly it measures with its so-called "grasp".

We conducted another test: we removed the correct answer to the Moutai question from the options list and the option "none of the above are correct", leaving only 3 descriptions that contradicted the facts. It almost entirely supported the option "revenue decline", but Moutai's revenue is actually increasing.

A more intuitive comparison is that the same error description is treated differently in the two tests.

When the option "None of the above are correct" is included, it provides almost no support for these descriptions; however, once the option is removed, the support rate jumps to over 80%, with no exceptions for any of the three companies.

TypeSafe explains this mechanism in its documentation: confidence measures "how much the model is biased towards one of these answers," it does not measure whether the answer itself is correct. A confidence level of 95% means "the most similar to this one among these," and it is incorrect to read it as "95% equals 100%."

It will only pick one from the answers you've written, and no matter how determined it is, it won't pick the one you didn't write.

IV. Derivative Value Flows

190 judgments, at the publicly listed price, actually cost less than 10 cents in total. One dollar could buy thousands of such judgments.

In the 19th century, after steam engines became more energy-efficient, Britain actually burned more coal because it became so cheap that it was worth using it in places where it wasn't previously worth burning coal. This is the starting point of Jevons' paradox.

TypeSafe incorporates this concept into the model name: for every order of magnitude reduction in the cost of an AI judgment, the increase in the number of calls will exceed the money saved. If the judgment can truly be converted into the price of a part, the software will generate a large number of call points that are not worth activating generative AI in the first place.

The application scenarios will most likely fall into two categories.

One approach is to layer armor on the large model workflow. This means that before each operation, the AI assistant determines whether a step requires human approval. The retrieved documents are first screened by the AI assistant before being fed to the large model. Some developer platforms have already released tutorials teaching people how to use this method.

Another category consists of a large number of trivial judgments that are not worth using AI for in the first place, including but not limited to work order assignment, refund reason classification, equipment alarm classification, product classification, and initial screening of contract terms. These are numerous, worthless per instance, and extremely sensitive to speed and cost.

However, we also discovered something during testing: the real cost is outside the interface bill.

In the challenging problems we deliberately selected, we set the threshold to the most conservative level, ensuring zero errors in the judgments that allow passage, and about 40% of the judgments were transferred to people.

If the threshold is lowered even slightly, mistakes will slip in; even if the threshold is raised to the highest level, it won't stop a list of options that has missed the correct answer.

Therefore, the derived commercial value will flow from the model in five directions, including the arithmetic rules and caliber processing written into the code, the option template designed by experts and with "none of the above are correct" exit points, the threshold repeatedly calibrated with their own business data, and the review team that takes over the transfer of traffic to manual review.

A working pipeline would generally look like this: the program is responsible for data retrieval and calculation, Jev is responsible for semantic judgment, and anything that is not certain is handed over to a human. The large model that can write is responsible for explaining the results to the human.

TypeSafe's chosen business strategy is very clear: integrate with multiple third-party platforms and gateways within a week, without developing its own application layer. It's betting on being integrated, running its software tens of thousands of times in other people's products, and charging based on usage.

However, there is an unverifiable risk: TypeSafe itself admits that it cannot prove that its current pricing is not dependent on subsidies.

A more profound impact on the tech industry may be the change in measurement systems. Over the past three years, the most common demand metric in the AI industry has been token consumption.

If high-frequency, low-complexity judgments are taken away from the general large model, this set of metrics anchored by "word count" will begin to distort.

The demand for inference computing power also varies: "long output, able to withstand a few seconds" and "short input, high concurrency, return within one second" require completely different hardware and scheduling structures.

V. However, the parts are not necessarily scarce.

A week after Jev's release, it seems that no domestic model company has yet made "judgment-only, no writing" into a separately priced commercial product.

However, China does not lack similar capabilities.

The LLaDA series developed by Ant Group and Renmin University of China also broke out of the framework of "writing one word at a time," but its purpose was to write faster. Jev simply didn't write at all, so the directions were slightly different.

The ability to make judgments exists extensively in China. The ranking models of Zhiyuan's BGE and Alibaba's Qwen3, as well as the content review and financial text tagging of various companies, run countless times every day in the search, risk control, and compliance links.

The key issue is that they are simply used as invoked models; no one packages them as an independent model category or sets an independent pricing unit.

In fact, open-source clones emerged just days after Jev's release: the most popular Laya uses the ModernBERT-large (421 million parameters) pure encoder architecture, while other projects directly fine-tune Alibaba Tongyi's open-source Qwen series models to run the same decision-making task.

The open-source model developed by the Chinese team became the foundation for its imitation, but the commercial product was actually created by a San Francisco company. Laya also stated in the project description that the calibration and training methods claimed by TypeSafe were not reproduced.

Calibration means that "when you say you're 80% sure, you're 80% right," which is the core difference that TypeSafe touts. The shell is easy to replicate, but how thick this layer is actually needs more independent testing to verify.

Our tests also yielded a finding that is more relevant to the domestic market:

For financial report chapters of the same length, the token consumption for Chinese materials is significantly higher than that for English materials, and it is also more likely to hit the limit for a single read.

Of course, TypeSafe itself admits that Chinese is weaker than English, and a native Chinese judgment model has room for improvement on paper.

The competitive narrative of domestic model companies is still locked on parameter scale, programming ability, and agent rankings. No one has yet taken the path of "driving the unit price of a single judgment to zero".

However, given the availability of underlying capabilities, the readily available open-source platform, and the natural alignment of application scenarios with China's massive financial data and content moderation needs, this business model is likely to be quickly adopted.

Whether it's fast or not depends on whether there's a team willing to turn it from a component into a business.

After three rounds of testing, we found that connecting to the model interface was actually easy, but the more difficult part was to thoroughly check the interim report data of the nine companies one by one, and then design a set of answers for each question without missing any options.

This business reduces the price of a single judgment to almost zero; the remaining price actually goes into the question setter's account.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.