CommunityJev Models & Reimplementations
Liquid AI d1-3B
Liquid AI's 3.12B decision model built on LFM2.5-VL: state plus typed questions in one forward pass with zero output tokens, 48.57 on Decision Index 0.2.1 and 8 ms per decision on an RTX 4090.
- Published by
- Community
- Author
- Liquid AI
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- One call reads the state once and answers every named question; the response's output_tokens is always zero.
- Score questions take 2–10 ordered levels with a legend — wider than the 0–5 scale other Jev-family models use.
- 8 ms per decision on an RTX 4090, 30 ms on an Apple M5 Pro, and 50 ms on a Jetson Orin Nano.
- Vision mean 74.1 across 11 public image benchmarks; removing the images drops the same questions to 45.1.
Watch out
The custom remote code needs transformers 5.14 or newer with trust_remote_code=True, the lfm1.0 license is not Apache, and the latency figures are warm-call numbers. It is not a chat model, and the generic vLLM or SGLang endpoints do not serve the decision interface.
Questions about this entry
Short answers drawn from the linked project or page.
No. It evaluates a state and typed questions in a single forward pass and returns answers with usage.output_tokens: 0; there is no generation path. Its Decision Index 0.2.1 score of 48.57 is billed as the best decision model under 10B, self-scored with the official kit.
The card lists an RTX 4090, an AMD MI325X, an Apple M5 Pro, Jetson AGX Thor, AGX Orin 64 GB, and Orin Nano. The compile(mode='reduce-overhead') path turns a single question into a CUDA graph that cut 16 ms to 8 ms on a 4090, with the first call per shape paying compilation.
Install transformers>=5.14 plus torch, torchvision, and pillow, then load with trust_remote_code=True. Call model.system_one(state, questions, images=None) for named questions in one pass, or system_one_batch for many requests packed without padding.
The lfm1.0 license, which is not Apache. The weights are LFM2.5-VL-3B post-trained for single-pass calibrated decisions, and they are marked custom_code, so they load with trust_remote_code=True.
Anywhere a pipeline needs a yes/no, a pick from named options, or a rating: routing and triage, moderation, intent and topic classification, extraction checks, reranking, LLM-as-a-judge scoring, agent guardrails, and visual inspection.
It scores 74.1 mean across 11 public image benchmarks, near the base model's 73.9, and with the images removed the same questions drop to 45.1, so the answers come from the pictures. The vision encoder is a SigLIP2 NaFlex 400M module.
Yes. One pass reads the state once and answers every named question, with output_tokens always zero, and system_one_batch packs many requests with no padding. With 64 states packed it reaches 262/s on a Jetson AGX Thor and 1,106/s on an AMD MI325X.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsHacker News: A one-token logprob wrapper that also classifies webcam frames.
allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker NewsMore Jev resources like this
Supersonic Labs' 144.3M-parameter decision model on mmBERT-small: choice, score, and noul answers from supplied options in one pass, 73.15% on 2,000 typed decisions, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
An untrained readout: stock Gemma-4-31B weights plus a decider config that reads option letters at the answer slot, scored #2 of 70 on Decision Index v0.2.1 at 57.33 against Jev 1.13.0's 57.91.
Jev Models & Reimplementations#community#open-models#local
Community
Nace.AI's 8B decision model on NVIDIA's Efficient-DLM diffusion backbone: a 256-dim pointer head scores every option in one pass, behind the same POST /v1/systemone API as its hosted service.
Jev Models & Reimplementations#community#open-models#local
Community
The vLLM Semantic Router project's 29.37B decision model on Qwen3.8-27B: a probability per supplied option with zero generated tokens, 74.0 on JevArena and 56.5 on Jev Decision Index 0.2.1.
Jev Models & Reimplementations#community#open-models#local
Community
A 4B scorer rebuilt from Qwen3-4B into a branching cross-encoder: one branch per option, no position bias, up to 256 options, and calibrated probabilities with no text to parse.
Jev Models & Reimplementations#community#open-models#local
Community
A 0.4B Jev-compatible decision model with measured order invariance: 960 random option shuffles moved no probability by more than 1e-5, and bev-decider serve exposes /v1/systemone.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
Interface, size, license, claims — in that order. A selection guide across the system-one catalog, from a 706K form scorer to 27B Jev students.
Jev Practices & Patterns#open-models#evaluation#benchmarks
Read article
Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.
Jev Models & Reimplementations#multimodal#open-models#decision-model
Read article
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
From the Jev community
Posts from builders shipping with Jev right now.