Dynamic Insight Beating AI News: Liquid AI has released the multimodal decision-making model d1-ultra, scoring 65 on Artificial Analysis's Decision Index v1.0, surpassing Jev 1.14's 57 points and ranking first.

The decision model does not handle chat or article generation; instead, it reads text and images directly to determine which option is more likely correct, providing a probability. For example, inspecting circuit board defects, verifying product images against descriptions, or reviewing insurance claim photos uploaded by customers.

However, there's an irony behind the new model's top ranking. In Artificial Analysis's 500-question knowledge test, d1-ultra correctly answered only about 32%, lower than the previous generation d1's 41%. Its strength lies in fewer incorrect guesses. Among questions it didn't answer correctly, the proportion choosing "don't know" rose from 2% to 85%. When uncertain, it prefers admitting ignorance rather than risking an incorrect response.

This also explains the significant rise in ranking. Artificial Analysis assigns 12.5% weight each to knowledge accuracy and avoiding wrong answers, with the rest of the score derived from complex reasoning, tool usage judgment, and safety classification.

Liquid AI's own evaluation of 112 tests shows d1-ultra leads Jev 1.14 across all eight task categories.

d1-ultra is now available via Liquid API and Vercel, but model weights are not yet open. It charges $0.042 per million input tokens—2.1 times that of the previous d1 generation—and no fee for output. If multiple questions are asked about the same image, each question incurs a separate image input cost.

Dynamic Insight AI News - The Ministry of Commerce and the Ministry of Science and Technology plan to include 23 AI technologies in export controls, covering large models, Agents, computing clusters, and more.

Among them, if a large model meets any one of the following three conditions, its core technology may be restricted:

1. Training computational requirements reach (10^{26}) floating-point operations

2. Parameter count reaches 2 trillion

3. Ranks in the top 10 globally in mainstream public evaluations

The controlled scope includes model weights, underlying source code, architecture design, and training optimization techniques.

However, this does not mean Chinese large models can no longer be open-sourced in the future. Open-source technologies already in the public domain are exempt from the new regulations; newly open-sourced restricted technologies can also go through a fast licensing procedure.

The Ministry of Commerce stated that it will generally achieve "apply and release immediately." Overseas users can continue using Chinese models via API, web pages, or apps, which are not within the scope of this large-model technology export control.

What is truly constrained is the transfer of core technologies overseas—for example, handing over model weights to foreign enterprises or deploying locally abroad.

Some technologies involving China's core data, national defense, and other sensitive areas are proposed to be prohibited from export.

The new list is currently under public consultation, with a deadline of November 15. However, the pre-export registration requirement for related technologies has already taken effect as of October 16.