AI BriefingLiquid AI Releases Open-Weight d1 Decision Models for Edge Inference
Key point
The new d1-3B model achieves a score of 48.57 on the Decision Index 0.2.1, outperforming larger models like Decider 35B-A3B.
Details
Liquid AI has released two open-weight decision models, d1-3B and the experimental d1-omni-600M, designed for fast, structured decision-making on edge devices. Unlike generative models that produce tokens, these models answer queries in a single forward pass, enabling low-latency inference for tasks such as intent classification, toxicity detection, and medical QA.
Performance and Benchmarks
The d1-3B model is positioned as the best decision model under 10B parameters on the Decision Index 0.2.1, scoring 48.57. This score surpasses all 4B and 9B models, including Decider 35B-A3B (47.11). On a suite of seven public datasets, d1-3B achieved a mean score of 82.9, exceeding Decider 4B (81.1). The smaller d1-omni-600M scored 78.4, outperforming Decider 2B (77.1) with only a quarter of the parameters.
| Benchmark | d1-omni-600M | d1-3B | Decider 2B | Decider 4B | | :--- | :--- | :--- | :--- | :--- | | SQuAD 2.0 | 74.0 | 83.3 | 67.7 | 76.0 | | Civil Comments | 95.8 | 93.3 | 93.6 | 92.8 | | MASSIVE intent | 86.1 | 86.9 | 81.1 | 88.3 | | PubMedQA | 61.3 | 68.3 | 65.7 | 63.3 | | BoolQ | 77.7 | 86.3 | 87.3 | 89.0 | | XNLI | 74.7 | 85.6 | 85.0 | 88.6 | | PAWS-X | 79.5 | 76.4 | 59.5 | 69.8 | | Mean | 78.4 | 82.9 | 77.1 | 81.1 |
Architecture and Modalities
The models are built on Liquid Foundation Models (LFMs) but utilize different backbones:
- d1-3B: Trained from LFM2.5-VL-3B, a decoder-only vision-language model. It supports text and image inputs.
- d1-omni-600M: Trained from LFM2.5-Encoder-350M, a bidirectional encoder. It supports text with either image or audio inputs. This model is currently in an early research release.
Edge and GPU Inference Speed
In collaboration with NVIDIA, Liquid AI evaluated d1-3B across various hardware. The model demonstrates sub-50ms latency for single-question responses on edge devices:
- NVIDIA Jetson AGX Thor: 16 ms per question.
- Jetson AGX Orin 64 GB: 26 ms per question.
- Jetson Orin Nano: 50 ms per question.
- Apple M5 Pro: 30 ms per question.
On high-end GPUs, latency drops further:
- NVIDIA RTX 4090: 8 ms per question.
- AMD MI325X: 9 ms per question.
The models are available on Hugging Face under open-weight licenses, requiring transformers>=5.14 for integration.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
TencentCloud Releases Octop, an Open-Source Self-Hosted Multi-Agent AI Assistant