The non-conversational AI has closed a financing round that values it at $7.5 billion, less than a month after its official launch.

The development company TypeSafe AI behind Jev recently completed $870 million in financing, with its valuation reaching $7.5 billion. This round of financing was led by Andreessen Horowitz, and Sequoia as well as existing investor DCVC also participated in the investment. At this point, only a few weeks have passed since the release of its new artificial intelligence model Jev.

At a time when chatbots are sweeping the globe, this "non-chatty" model has instead become the target pursued by both capital and enterprises. TypeSafe claims that one-third of Fortune 500 companies are already using Jev, and the speed of enterprise adoption is remarkable. More critically, Jev is not a large language model. It does not output text, but generates probabilities — in TypeSafe's words, it delivers "calibrated decisions".

Became a hit right after release: The unexpected rise of a "mute model"

On September 15, Jev became popular almost instantly after going online. A large number of usage scenarios quickly emerged in the developer community: some used it to filter spam information, some used it to assist trading decisions, and others connected it to the Agent system to act as a "referee". The release video received tens of millions of views on social platforms, and the number of paid teams adopting it set a new record on the cloud development platform.

The reaction of the capital market is equally rapid. When TypeSafe AI just came out of stealth mode, it completed a $40 million seed round of financing led by DCVC, with a valuation of about $200 million. However, the explosive growth of Jev suddenly accelerated the financing pace. According to reports from *The Information*, the company was once negotiating for at least $1 billion in new financing, and some investors offered a valuation quote of more than $10 billion. In the end, the $870 million financing led by a16z was closed, and the valuation was fixed at $7.5 billion.

TypeSafe AI was co-founded in 2024 by Sasha Sheng, a former Meta research engineer, and Erik Gafni, an engineer and entrepreneur. Almeida previously served as a researcher at OpenAI. The team is not large, but it has reached a valuation of $7.5 billion in a short period of time. The outside world still knows limited details about the founding team, but capital obviously cares more about what Jev itself can do.

Instead of generating text, it generates "decisions". Jev is based on the Transformer architecture, but it is not a large language model. It does not output text, but generates probabilities, or "calibrated decisions". The reason why Jev excites users and large enterprises is that TypeSafe claims it runs much faster than LLMs and uses far fewer tokens. The company positions its approach as particularly suitable for automating tasks, rather than generating text or code.

"For four years, we have done a very good job with human language, but this is useless for automation, because computers speak a different language," Diogo Almeida, co-founder of TypeSafe, told TechCrunch last month.

This sentence points out the core logic of Jev. Traditional large language models output natural language. If applications want to use these results, they often need to write additional code to parse the text and extract the values that are really needed — such as a classification label, a yes/no judgment, or a score. This parsing step not only increases development complexity, but also introduces uncertainty.

Jev takes a completely different approach. It classifies input information into a preset output set, and directly returns "yes or no" judgments, numerical scores, or results of established options. It returns typed values that applications can use directly, without intermediate parsing steps. In other words, Jev is not designed to be understood by humans, but to be called directly by code.

This set of technical ideas is called "reinforcement learning for calibrated decisions" by TypeSafe. Almeida previously participated in the construction of instruction-following models at OpenAI and was one of the co-inventors of RLHF technology. But it was precisely the questioning of RLHF that prompted him to leave OpenAI. He revealed a key insight in a podcast interview: although RLHF training makes the model better at conversations, it also quietly destroys the model's calibration ability, leading to "mode discard". In other words, the price of making the model learn to speak beautifully is that it loses the ability to make reliable judgments.

Jev's positioning is therefore particularly clear: it is not designed to chat with people, but to be called by code. Its output response time is within 700 milliseconds, which is about 100 times faster than cutting-edge large language models, and its cost is only 1/100 to 1/500 of the latter. Official data shows that when processing the same batch of comment labeling tasks, Jev takes 203.2 seconds and costs $0.84, while the comparison model takes 823.5 seconds and costs $1.50. The current pricing of Jev is only $0.042 per million input tokens, with free outputs. a16z even used a detail in its investment announcement to illustrate the compactness of Jev's outputs: "The outputs are so small that they don't even charge for them."

Why are enterprises willing to pay for it?

There has long been a pain point in automation scenarios: although large models are smart, they are not stable enough. Almeida once used a metaphor to describe this dilemma: if a model "is as smart as Einstein 95% of the time, but behaves like Mr. Bean for the remaining 5%", then it is useless in automated production environments.

What Jev tries to solve is exactly this problem. It does not pursue to generate fluent text, but to make reliable judgments within limited options. This capability is exactly the rigid demand in high-frequency scenarios such as classification and screening, process approval, path selection, and agent action judgment. Enterprises do not need an AI that can write poetry, but an AI that can tell the system "pass" or "reject" within 700 milliseconds.

The trend of industry division of labor is taking shape: generative large models are responsible for complex reasoning, and decision models are responsible for high-certainty automated judgments. Almeida calls this direction "System One models", which are designed to be deeply integrated with code rather than to talk with humans.

a16z elaborated on the far-reaching significance of this concept in its investment announcement: "Ironically, the phrase 'AI eats software' means that it is writing more traditional software faster, but not necessarily making software smarter. The real economic unlock should be that all software can natively embed the intelligence of the latest models with performance and cost that conform to the economics of traditional software."

Giants follow up, and a new track takes shape

The sudden popularity of Jev quickly triggered a chain reaction among Silicon Valley giants.

OpenAI launched a tool called Decisions API, which relies on its Luna model to answer questions limited to preset results. Data platform Databricks released the ai_decide feature. Cloudflare launched the open-source Clef series. Amazon previously had Strands Decider 2B launched and put into use.

These products point to the same direction: instead of adopting the token-by-token autoregressive generation mode, they directly score candidate answers, targeting high-frequency limited-option scenarios such as classification and screening, process approval, path selection, and agent action judgment. Jev is not the only player, but it is the first model to bring this route into the spotlight.

Almeida said that he had long expected that copycat competitors would appear, but he still believes that Jev has stronger capabilities and higher level of intelligence. "Jev represents the beginning of a new type of artificial intelligence," he said. "There are still a large number of untapped technical directions in the field of artificial intelligence, and Jev has only taken the first step in this new track."

Why is capital going crazy?

TypeSafe AI took less than a month to grow its valuation from $200 million to $7.5 billion. This rate of valuation surge is rare even in the history of AI financing.

a16z did not hide its excitement in the official investment announcement: "The release of Jev is the biggest narrative disruption we have seen this year." The top venture capital firm, which has invested in OpenAI and xAI, wrote: "We have never seen a model grow so fast — it reached 1 trillion tokens of generation volume just 3 days after going online."

The valuation logic is not complicated. Jev does not try to replace ChatGPT, but embeds into software automation processes, meeting the performance and cost requirements of traditional software economics. If enterprises can embed intelligent decision-making capabilities into software at extremely low cost and extremely high speed, its market space will not be limited to AI conversations, but the entire software automation market.

Of course, challenges also exist. Copycat competitors are emerging, and large model vendors may also internalize similar capabilities into their own products. Enterprises are adopting it at a surprising speed, but long-term retention and reliability still need to be verified. The performance of calibrated decisions in terms of safety, responsibility and boundary conditions will also be subject to stricter scrutiny.

But at least for now, the market has given its judgment with real money: in automation scenarios, reliability is more important than intelligence, and executability is more valuable than readability. When the entire industry is pursuing AI that is more like humans, Jev has chosen a path that "behaves more like a computer". And the commercial value of this path is being rapidly revalued.

This article is from the WeChat Official Account "Tech Business", authored by Tech Business, and published with authorization by 36Kr.