Jev vs d1-3B
Open 3B decision model for text and images, aimed at edge hardware.
Jev 1.13.0
TypeSafe · hosted System One model
77.1Capability · unranked reference
Intelligence 63.6 · calibration 90.6.
d1-3B
Liquid AI · Open 3B decision model for text and images, aimed at edge hardware
—Decision Index v0.2.1 · 48.57
Decision Index v0.2.1 public score 48.57. Not a JevBench v1.6.1 rank.
Key differences
- Where it runs JevHosted APId1-3BSelf-hosted, including Jetson-class boards
- Inputs JevText · images (experimental, 1.14)d1-3BText and images in one state
- How it answers JevTyped probabilities, no essayd1-3BOne forward pass, zero output tokens
- Published score JevJevBench v1.6.1 Capability 77.1d1-3BDecision Index v0.2.1 public set 48.57
- Size JevHosted; weights are not publicd1-3B3B, from LFM2.5-VL-3B
- Sibling Jev—d1-3Bd1-omni-600M, experimental, text+image or text+audio, index 15.95
Do not mix the two scoreboards
Liquid published d1 on the Decision Index. Benchmark Heaven’s October Capability list, which this site uses for the other new pages, does not include a d1 row. Jev’s 77.1 is not a score on Liquid’s index. Model card ↗
Numbers Liquid published
Decision Index v0.2.1, public set: d1-3B 48.57, d1-omni-600M 15.95, Winnow-12B 50.02, Decider 35B-A3B 47.11.
Seven public text benchmarks, mean: d1-3B 82.9, d1-omni-600M 78.4. Liquid puts Decider 2B at 77.1 on that same mean — that 77.1 is not Jev’s JevBench score.
Latency for one d1-3B question: 8 ms on an RTX 4090, 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin, 50 ms on Jetson Orin Nano.
What d1-3B is
d1-3B is Liquid AI’s open decision model, released 8 October 2026 with the experimental d1-omni-600M. You give it a state — text, JSON, images, or a mix — and a set of questions. It returns calibrated typed answers in one forward pass and does not generate output tokens. The 3B checkpoint averages LFM2.5-2.6B and LFM2.5-VL-3B text weights, then fine-tunes and merges them.
Liquid’s Decision Index v0.2.1 public score for d1-3B is 48.57. On that table it sits just behind Winnow-12B (50.02) and ahead of Decider 35B-A3B (47.11). d1-omni-600M scores 15.95 on the same index. Jev is not in that published table, and d1 is not in the JevBench v1.6.1 alternatives list this site uses. Those two numbers are not a head-to-head.
Read this before you switch
- Decision Index 48.57 and JevBench Capability 77.1 are different tests. Neither number says which model wins on your tickets.
- d1-omni-600M is an early research checkpoint. Liquid says it is still in development, and they did not publish its latency.
- Liquid’s 8 ms figure is one question on an RTX 4090. The same card quotes 16 ms on Jetson AGX Thor, 26 ms on AGX Orin, and 50 ms on Orin Nano.
When to use which
Choose Jev if
- You want a hosted text API with a JevBench v1.6.1 reference score, and no device to flash.
- Your volume is server-side classification, routing, or evaluation — not a 3B model on a Jetson.
- You need the request shape and versioning you already use with Jev.
Choose d1-3B if
- The decision has to run on your own machine — Jev is hosted only (its image input is experimental and cloud-side).
- You want a 3B open checkpoint with llama.cpp support and a single-digit-millisecond path on a 4090.
- You can evaluate d1 on your own labels instead of treating Decision Index as a JevBench rank.
FAQ
Is d1-3B better than Jev?
There is no shared public rank to answer that. d1-3B scores 48.57 on Liquid’s Decision Index v0.2.1. Jev’s figure on this site is JevBench v1.6.1 Capability 77.1. d1 is not on that alternatives list, and Jev is not in Liquid’s index table.
What is d1-omni-600M?
Liquid’s experimental sibling. It starts from LFM2.5-Encoder-350M and adds vision and audio. A request is text plus image, or text plus audio. Its Decision Index v0.2.1 score is 15.95.
Does d1 generate text?
No. Liquid describes one forward pass and zero output tokens: typed, calibrated answers only. That part of the job matches Jev.