d1-2B: What's New, Hardware Needed, How to Run It Free

d1-2B decides on text, images, and video in one forward pass, without generating tokens. It can return an answer in just 5.7 ms on an NVIDIA H100.

According to LiquidAI, d1-2B is a decision-oriented model that works with text, images and video. It uses the LFM3-VL-2B backbone and has 1.92 billion parameters. It can answer in multiple languages, including English, Arabic, Chinese and many others. The model returns typed answers with probabilities in a single forward pass, so there are no sampling parameters to set.

The model card says d1-2B returns calibrated answers in one forward pass, unlike typical generative models that emit tokens. You can ask any number of named questions about a single state, and the model will answer them all at once. Beyond simple classification, it can extract text, numbers, dates with spans, locate objects in pictures, and even give quantity estimates with confidence intervals. All of this happens without generating any output tokens.

LiquidAI reports that d1-2B leads the open d1 family on average across a board of more than a hundred decision benchmarks. On the Decision Index public suite it scores 51.1, and on the overall board it reaches 63.2, ahead of the larger d1-3B model. In vision tasks it scores 76.7 on AI2D and 86.1 on CV-Bench, placing it among the top open models for image decisions.

The model runs in about 5.7 ms per decision on an NVIDIA H100 and 3.8 ms on an AMD MI325X, making it suitable for high-throughput pipelines. It excels at tasks like routing, moderation, intent classification, document extraction and visual inspection. However, the card notes that knowledge-heavy queries favor larger models and accuracy drops after about 64 K tokens of state. Also, d1-2B is a decision model, not a conversational chat model.

If you run AI at home, the model fits on a modest RTX 4060, according to the data. It is ideal for developers who need rapid, calibrated decisions over text, images or video without the overhead of token generation. Projects that combine multimodal inputs, like forms with attached pictures, can benefit from its unified interface. And anyone who prefers a decision output over a generative response will find d1-2B a good match.

What do you need to run it at home? At 4-bit it takes about 4 gigabytes of memory. The smallest home machine that holds it fully is an RTX 4060. On an RTX 3060 12 GB with 32 gigabytes of RAM it writes about 183 tokens per second. On an RTX 4090 with 64 gigabytes of RAM it writes about 385 tokens per second.

There is no free API for it yet. The weights are open: you can download them from Hugging Face, under the license written in its model card. Arena has not ranked it yet: we add the score as soon as the votes arrive.

Compare d1 2B with every free model, and check what your own GPU can run, on FreeLLMArena.