Article

Liquid AI decision model upgrade: 5 new capabilities added, tops the ranking with 75 points, price unchanged

Dongcha Beating AI news flash: Liquid AI has released two multimodal decision models, d1.2 and d1.2-ultra.

The new models scored 69 points and 75 points respectively on the Artificial Analysis decision model leaderboard, up 3 points and 4 points from their respective predecessors. Among them, d1.2-ultra ranks first, leading second place by 2 points.

d1.2-ultra's context window has also been expanded from 33000 to 66000 Token. Average task time has been shortened from the previous generation's 301 milliseconds to 218 milliseconds, a reduction of about 28%.

Previously, d1-ultra mainly handled yes/no judgments, single-choice selection and scoring; the new version adds 5 categories of tasks: extracting specific values such as amounts and dates from documents; finding all text that meets given conditions along with its position; circling specified objects in images; estimating quantities and giving a range; selecting multiple qualifying options at once.

For example, when processing invoices, the model can now directly extract the amount payable and return the corresponding original text. If nothing is found, it answers "not found." When processing images, it can also box all objects matching a description, rather than just judging whether they exist.

Both models also add audio input, able to directly recognize sounds such as languages and sirens in recordings, and also supporting asking questions by voice.

Liquid AI's internal tests show that d1.2-ultra's accuracy in extracting values from receipts reaches 98%, and image target localization reaches 94%.

In another comparative test covering 22,505 questions, its confidence calibration error was only 0.8 percentage points, while Perplexity Decider v1.3's was 3.7 percentage points. In other words, the judgment probabilities it gives are closer to its actual accuracy rate, making it easier for systems to identify tasks that need human review.

Prices remain unchanged. d1.2 and d1.2-ultra charge 0.02 USD and 0.042 USD respectively per million input Token, with no output fees, and both APIs are now open.

Liquid AI says it will release an open-source version with comparable capabilities in the future.