The decision-model lane has been crowded lately: within two weeks Microsoft, Amazon, Cloudflare and Perplexity all open-sourced their own models, with an identical playbook — compress LLM tool-call routing into one cheap classification. What Liquid AI just shipped is different: multimodal, per-frame, zero-token output.

A model that generates no tokens

d1-3B and d1-omni-600M belong to Liquid AI's d1 decision-model family — 3.1B and 600M parameters, post-trained from two different backbones, LFM2.5-VL-3B and LFM2.5-Encoder-350M. The fundamental difference from a generative LLM: it emits no tokens. You hand it a state (text, JSON, an image, or a mix) plus a set of questions, and a single forward pass returns calibrated, typed answers — yes/no, a choice among named options, a rating level. The official API's usage field reports output_tokens as exactly 0.

That design pushes latency into territory generative models cannot reach: 8 ms per single question on an RTX 4090, 9 ms on an AMD MI325X, 16 ms on a Jetson AGX Thor, and 50 ms even on the smallest Jetson Orin Nano. Packing three questions into one forward pass moves the AGX Thor from 16 ms to just 20 ms; with 64 states packed together, the AGX Thor processes 262 per second. In collaboration with NVIDIA, there is also a hardware-in-the-loop demo in Isaac Sim — the model runs on a Jetson board and makes navigation decisions directly.

Benchmarks: a 3B model ranked second overall

On the public Decision Index v0.2.1 leaderboard, d1-3B scores 48.57 — second only to the 12B Winnow-12B (50.02), ahead of every other sub-10B model, and above the 36B Decider 35B-A3B (47.11), a model roughly 12x its size. In the sub-scores, Tools at 74.5 and Arts at 36.3 are the highest on the board.

Across seven public text benchmarks (SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI, PAWS-X), d1-3B averages 82.9, above Decider 4B (81.1) and Decider 2B (77.1). The smaller d1-omni-600M averages 78.4, beating Decider 2B with about a quarter of the parameters; its toxicity-detection score (Civil Comments 95.8) and paraphrase-identification score (PAWS-X 79.5) are the highest in the table.

Training post-mortem: plain tricks won

The most interesting line in the official write-up: training on long inputs, shuffling answer options, and fixing shortcuts in the data "made a bigger difference than more advanced techniques." The d1-3B base itself comes from directly averaging the weights of LFM2.5-2.6B and the text backbone of LFM2.5-VL-3B, then fine-tuning multiple checkpoints with different random seeds and data mixtures before merging them again. The 600M omni variant added modalities in stages: audio via a FastConformer encoder with an adapter, vision borrowed from the LFM2.5-VL-450M encoder with LoRA updates — active only when the input includes images, with the vision encoder frozen throughout.

Multimodal is the differentiator, and the gaps are stated plainly

Across 11 public vision benchmarks, d1-3B averages 74.1 versus the backbone's 73.9 — essentially flat: MMVP rises from 73.7 to 77.0 and VL-RewardBench jumps from 50.9 to 65.0, but CV-Bench falls from 87.6 to 82.1 and VisualWebBench from 78.3 to 71.4. The company published both directions in the same table. With the images removed, the same questions drop to 45.1, showing the answers genuinely come from the images. The omni-600M is a rare text+image+audio triple-modality combination at the 600M scale, but its 15.95 Decision Index score is explicitly labeled an experimental early checkpoint, and dedicated audio decision benchmarks "are currently an open problem" — a gap left to the community.

All 10 companion demos run on a Hugging Face Space, calling d1-3B frame by frame from a live camera: gesture-controlled games, real-time content moderation, one answer read per frame. On the inference side, llama.cpp is supported from day one across Apple, AMD, Qualcomm and NVIDIA (including NVFP4), and the weights are open for download.

So what

Five players — Microsoft, Amazon, Cloudflare, Perplexity and now Liquid AI — have bet on decision models within two weeks, which says that "splitting routing and classification out of the LLM" is now industry consensus. But the first four all frame it around text-routing cost-performance and leaderboard scores; the real increment from d1-3B is making "you can ask the model once per frame" an engineering reality — a 50 ms Orin Nano means a decision model can sit inside a camera pipeline working frame by frame, something a generative LLM structurally cannot do. Trading 3B for a 12B-wide gap to first place looks fine anywhere edge deployment matters. The next thing to watch is whether audio decision benchmarks emerge — that is where the omni variant would truly belong.

Sources: Liquid AI blog · d1-3B model card