It’s something that nobody in the AI space really expected: after GPT-Astra and Fable were released, there’s still been such a hot new type of model that has caught fire.
JEV’s popularity comes from two sides. On the one hand, it genuinely solves certain problems. The other key reason is that its speed advantage is extremely easy to spread.
It has both speed and scale. You’ll find that the demo cases on those webpages all highlight its speed and the volume it can handle. That’s the most immediate, visceral kind of appeal—and it also shows that everyone really cares about these two aspects, which is why all kinds of demos and case studies are being shared like crazy.
Of course, this model isn’t万能 either. If you force it to work on its own, without relying on large language models, the application scenarios are actually quite limited. But if you add assistance from a large model or other tools, the results can be excellent.
At the same time, it still lacks multimodality. If it had multimodality—especially for organizing and categorizing materials or for producing long-form content—the experience would be even more satisfying.
Just yesterday, they announced it’s open to everyone now, and there’s no need for a waitlist. Each user gets a $5 free quota.
This model is really cost-effective: input costs just $0.042, and the output is free—so that $5 can last a very, very long time.
And coincidentally, we updated Next Token last week, mainly discussing:
The JEV model and related issues—for example, people can train models similar to it within a day. Big players like GPT-6 and Fable are already extremely professional in model training, which leaves plenty of room for “model training arbitrage” with smaller models.
The collaboration between small models and large models.
The evolution of the interaction interface: now chatting, browsing, and task lists are all in the same piece of software. This includes Codex’s recent browser plugin update—it also supports multiple accounts (for example, Notion’s MCP can log into multiple accounts and pull information). In many cases, it’s already replacing our system interface. This new kind of interaction may, almost imperceptibly, also educate users.
Yesterday, I also used JEV to create a real-time 3D scene generation demo. People who liked it were quite a lot. In the show, we discussed these as well as other news items, and we also shared our views—me, Qiao Mu, Yang Pan, and Juzi. Feel free to go listen.
JEV’s popularity comes from two sides. On the one hand, it genuinely solves certain problems. The other key reason is that its speed advantage is extremely easy to spread.
It has both speed and scale. You’ll find that the demo cases on those webpages all highlight its speed and the volume it can handle. That’s the most immediate, visceral kind of appeal—and it also shows that everyone really cares about these two aspects, which is why all kinds of demos and case studies are being shared like crazy.
Of course, this model isn’t万能 either. If you force it to work on its own, without relying on large language models, the application scenarios are actually quite limited. But if you add assistance from a large model or other tools, the results can be excellent.
At the same time, it still lacks multimodality. If it had multimodality—especially for organizing and categorizing materials or for producing long-form content—the experience would be even more satisfying.
Just yesterday, they announced it’s open to everyone now, and there’s no need for a waitlist. Each user gets a $5 free quota.
This model is really cost-effective: input costs just $0.042, and the output is free—so that $5 can last a very, very long time.
And coincidentally, we updated Next Token last week, mainly discussing:
The JEV model and related issues—for example, people can train models similar to it within a day. Big players like GPT-6 and Fable are already extremely professional in model training, which leaves plenty of room for “model training arbitrage” with smaller models.
The collaboration between small models and large models.
The evolution of the interaction interface: now chatting, browsing, and task lists are all in the same piece of software. This includes Codex’s recent browser plugin update—it also supports multiple accounts (for example, Notion’s MCP can log into multiple accounts and pull information). In many cases, it’s already replacing our system interface. This new kind of interaction may, almost imperceptibly, also educate users.
Yesterday, I also used JEV to create a real-time 3D scene generation demo. People who liked it were quite a lot. In the show, we discussed these as well as other news items, and we also shared our views—me, Qiao Mu, Yang Pan, and Juzi. Feel free to go listen.
