Skip to content
Bud's Wiki

Open d1: Edge decision models for text, vision, and audio

TL;DR

Section titled “TL;DR”

Liquid AI releases Open d1: two open-weight multimodal decision models in its d1 family. d1-3B (built from LFM2.5-VL-3B) handles text + images and tops the Decision Index v0.2.1 public split for sub-10B models, matching a 35B-A3B decider at 1/12 the parameters. d1-omni-600M (built from LFM2.5-Encoder-350M + vision + audio adapters) adds text+image and text+audio in a single bidirectional encoder. Both models produce an answer in a single forward pass rather than generating tokens, with day-one llama.cpp support and reported edge latencies of 8 ms on an RTX 4090, 16 ms on Jetson AGX Thor, and 50 ms on Jetson Orin Nano for a single question. This is Liquid AI’s open-weights entry into the “System One / decision model” class pioneered by TypeSafe AI’s closed Jev and open-sourced by Convai’s Laya.

Key claims

Section titled “Key claims”
  • Decision models in the d1 family don’t generate tokens; they produce an answer in a single forward pass [§Architecture and Training].
  • d1-3B scores 48.57 on the Decision Index v0.2.1 public split, ahead of every model under 10B and on par with Decider 35B-A3B (12× its size) [§introduction].
  • d1-3B is trained from LFM2.5-VL-3B (decoder-only VLM), via weight-averaging the LFM2.5-2.6B and the LFM2.5-VL-3B text backbone, then fine-tuning with different seeds + data mixtures and merging again [§Architecture and Training].
  • d1-omni-600M is trained from LFM2.5-Encoder-350M (bidirectional encoder) with staged addition of FastConformer audio and LFM2.5-VL-450M vision encoders, LoRA-adapted only when image input is present [§Architecture and Training].
  • Mean text-benchmark score of 82.9 for d1-3B across SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI, PAWS-X — ahead of Decider 4B (81.1) [§Benchmarks].
  • d1-omni-600M mean 78.4 outperforms Decider 2B (77.1) at a quarter of the parameters and tops the table on toxicity detection (Civil Comments 95.8) and paraphrase identification (PAWS-X 79.5) [§Benchmarks].
  • Single-question latency for d1-3B: 8 ms on RTX 4090, 9 ms on AMD MI325X, 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin 64 GB, 50 ms on Jetson Orin Nano [§Fast Inference Everywhere, §GPU inference].
  • Scaling to three questions over the same state takes only ~1.3× the time of one (AGX Thor: 16 ms → 20 ms), reflecting the shared-prefill structure [§Edge inference].
  • 384-px image inference is 17–18 ms on RTX 4090 / MI325X and 35–202 ms across the Jetson lineup [§Fast Inference Everywhere].
  • A dedicated audio-decision benchmark is called out as an open problem — d1-omni-600M’s audio capability is demonstrated via playground demos rather than scored numerically [§Benchmarks].
  • Both models ship with day-one llama.cpp support and NVFP4 across Apple, AMD, Qualcomm, and NVIDIA hardware [§Get Started].

Method

Section titled “Method”

d1-3B and d1-omni-600M are built from two different Liquid Foundation Model backbones and converted into single-forward-pass decision heads. For d1-3B, Liquid first averages the weights of LFM2.5-2.6B and the text backbone of LFM2.5-VL-3B to form a better starting checkpoint, then fine-tunes multiple variants with different random seeds and data mixtures and merges them again. The reported heuristic is that training on long inputs, shuffling answer options, and fixing data shortcuts mattered more than any single algorithmic trick.

For d1-omni-600M, the recipe is more surgical: start from the bidirectional LFM2.5-Encoder-350M, fine-tune on decision tasks, then add modalities in stages. A FastConformer audio encoder with an adapter is trained first (text backbone frozen), then adapted with a frozen text backbone. The vision stage reuses the encoder from LFM2.5-VL-450M and adds an adapter plus LoRA updates to the backbone — the LoRA is active only when the input contains an image. A final full-model fine-tune, LoRA merge, and weight-average with the previous checkpoint regularizes the result. The architecture detail that makes this cheap is that neither model emits tokens: the schema-conditioned decision is read out from fixed positions in a single pass, so audio and vision streams only need to feed into the shared encoder prefix.

Results

Section titled “Results”

On the Decision Index v0.2.1 public split, d1-3B scores 48.57 and d1-omni-600M scores 15.95; the private vision split is held back for the Decision Index v0.3. Across seven text benchmarks d1-3B averages 82.9 (vs 81.1 for Decider 4B) and d1-omni-600M averages 78.4 (vs 77.1 for Decider 2B at ~4× the parameter count). Highlighted per-benchmark wins: Civil Comments 95.8 (toxicity), PAWS-X 79.5 (paraphrase), MASSIVE intent 87.3, PubMedQA 65.7.

Latency results, measured one request at a time: d1-3B answers a single question in 8 ms on RTX 4090, 9 ms on MI325X, 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin 64 GB, 50 ms on Jetson Orin Nano. A 384-px image is processed in 17 ms (RTX 4090) to 202 ms (Orin Nano). Sustained question throughput reaches 1,106/s on MI325X and 475/s on RTX 4090. Three structured questions over one state cost only ~1.3× a single question — the shared-prefill design pays off exactly as the model-class promises.

Why it’s interesting

Section titled “Why it’s interesting”

Open d1 is the third major entry in the System One / Decision Models cluster. Introducing System One Models & Jev launched the category as a closed-source product with no technical report and no public benchmarks, framing it as a new model class trained with Reinforcement Learning for Calibrated Decisions. Laya — Non-Autoregressive System One Decision Model (open-source RLCD) gave the field its first serious open-weights rebuttal — a 421M ModernBERT-large + option-marker head trained with a strictly proper scoring rule, beating Jev’s reported numbers on third-party benchmarks at 7.8× lower latency. d1-3B now pushes the open side up to the 3B / multimodal frontier: it keeps the single-forward-pass property but adopts a decoder-only VLM backbone (LFM2.5-VL-3B) rather than Laya’s bidirectional encoder, and reports hardware-in-the-loop demonstrations with Isaac Sim on Jetson — the first filed instance of a System One model wired into an embodied-AI control loop.

The multimodal footprint is the sharpest differentiator from Laya. d1-omni-600M is the first filed decision model that handles text+image or text+audio in a single checkpoint, pushing the model class past Jev’s and Laya’s text-primary origin point. Open-weights releases that bundle vision + audio + text have historically landed on Open foundation-model releases, and this one adds a specifically decision-model-shaped twist on the pattern. The flip side is that the Decision Index v0.2.1 private vision split isn’t scored and audio decision benchmarks are explicitly called an open problem — the audio capability is a playground demo at launch, not a benchmarked result.

The three-way comparison also sharpens a question raised but not resolved by the earlier two filings: how much of the “new model class” story is architecture vs. task-shaped fine-tuning. Laya’s typed-decisions win required fine-tuning on the benchmark’s own training split, and d1-3B’s benchmark win is reported against other dedicated deciders rather than against the same backbone (LFM2.5-VL-3B) prompted zero-shot into structured output. A fourth datapoint that holds backbone fixed and ablates the single-pass training recipe would answer this cleanly.

See also

Section titled “See also”