Liquid AI releases d1 decision model and open-source Pipette on-device benchmark

Liquid AI launched d1, a model that returns typed decisions without generating tokens, and Pipette, an open-source suite for on-device model benchmarks.

Liquid AI has released d1, a "decision model" for classification, routing, approval gates and scoring. Alongside it comes Pipette, an open-source suite for measuring model quality, speed, latency and memory on phones and laptops, built in partnership with Artificial Analysis (Liquid AI thread). If you use a general LLM just to emit a label or a boolean in JSON, d1 targets that job. Pipette gives you a way to check a quantized model on real hardware before you ship it.

According to AlphaSignal, each d1 call takes a state and one or more questions. Each question uses one of three answer types: Noul (yes/no probability), Choice (a distribution over options) or Score (a position on an ordered rubric). The model reports zero output tokens. A single request can mix these types, for example intent, urgency and bug probability for one support ticket. It is available now through the Liquid API with a d1:free tier. OpenRouter support is only "coming soon," and we couldn't find paid pricing. Liquid calls the probabilities calibrated. AlphaSignal advises checking calibration on your own traffic first.

Liquid says d1 is the first model to beat Jev on a Decision Index it attributes to Hugging Face. AlphaSignal reports that Jev 1.13 led the index with a composite near 74.4 before d1 took the top spot; we couldn't find d1's own score from a reliable source. It is unclear who runs the index or how independent it is.

Pipette compares configurations defined by model, quantization, runtime and device. Liquid says its launch dataset holds 10k+ results across about 35 model classes and 7 llama.cpp quantization levels. Hardware coverage is narrow for now: an M5 Max MacBook Pro, iPhone 17 Pro and Galaxy S26 Ultra, with a Ryzen AI Max+ 395 listed as coming soon. Community submissions are planned.

Sources