Liquid AI has introduced Open d1, a pair of open-weight multimodal models named d1-3B and d1-omni-600M, expanding its d1 decision model collection. The d1-3B model is proficient in processing both text and images, while d1-omni-600M can interpret text alongside either an image or an audio input. A key innovation of both models is their ability to deliver calibrated, structured responses in a single forward pass, generating zero output tokens. This design aims for ultra-fast, real-time decision-making, especially on NVIDIA's hardware ecosystem, including DGX servers, RTX workstations, and Jetson edge boards.
Accelerated Deployment and Accessibility
Both models are designed for immediate deployment. Their checkpoints are readily available on Hugging Face, compatible with the popular Transformers library, and boast day-one support for llama.cpp. Under the LFM Open License v1.0, commercial use is permitted free of charge for entities with annual revenues below $10 million. It is worth noting that d1-omni-600M is presented as an early research release, with specific latency figures yet to be published.
Understanding Decision Models
In contrast to traditional generative large language models (LLMs) that construct answers token by token for subsequent parsing, a decision model operates differently. It receives an input state and a predefined set of questions, then processes them once to return a probability distribution for every permissible answer. Liquid AI defines three distinct question formats:
- Noul: A binary yes or no inquiry, yielding a probability for 'yes'.
- Choice: Requires selecting a single label from a list of named options, providing a full probability distribution across them.
- Score: Assigns a probability-weighted position along an ordered rubric, typically featuring between 2 and 10 levels.
Multiple questions can be processed concurrently from a single input state in one call, consistently reporting zero output tokens. Liquid AI recommends d1 models for diverse applications, including routing, content moderation, intent classification, re-ranking, LLM-as-a-judge scoring, agent guardrails, and visual inspection. These models are not built for conversational chat.
Architectural Insights
The d1-3B model comprises 3.12 billion parameters and is based on LFM2.5-VL-3B, a decoder-only vision-language model. Its development involved averaging the weights of LFM2.5-2.6B with the text backbone of the base model, followed by fine-tuning with various checkpoints and data mixtures before another merging phase. It incorporates a 400-million parameter SigLIP2 NaFlex vision encoder and supports a 32,768-token context. Engineering emphasized the importance of handling long inputs, shuffled answer options, and addressing data shortcuts over more complex algorithmic techniques.
The d1-omni-600M model, with 587 million parameters, originates from LFM2.5-Encoder-350M, a bidirectional encoder. Its architecture features a 381-million parameter shared core and decision head, a 94-million parameter vision encoder, and a 112-million parameter audio encoder, which is a 17-layer FastConformer. The model handles a 16,384-token context, with audio clips capped at 30 seconds. It processes either images or audio alongside text, but not simultaneously. Audio training specifically targeted English speaker-to-assistant requests.
Optimal Use Cases
The Open d1 models are particularly well-suited for:
- Support and Ticket Triage: A single d1-3B call can evaluate refund eligibility, assign tasks to specific teams, and rate message urgency from the same input, all with no output tokens to parse.
- Real-time Visual Inspection and Moderation: At the edge, d1-3B processes a 384px image in approximately 35 milliseconds on a Jetson AGX Thor. Liquid AI's Open d1 Arcade demonstrates its capability to run frame-by-frame on live camera feeds for content moderation and gesture control.
- Voice-Command Routing: On compact devices, d1-omni-600M can process up to 30 seconds of speech alongside text, directly identifying the speaker's intent or topic for immediate routing.
Benchmark Performance
On Decision Index v0.2.1, d1-3B achieved a score of 48.57, outperforming all other models under 10 billion parameters and slightly exceeding Decider 35B-A3B (47.11). Only Winnow-12B scored higher at 50.02. Liquid AI conducted these benchmark tests internally. The d1-3B model led in the Tools (74.5) and Arts (36.3) categories but showed lower performance in Knowledge (23.8).
Across seven public text benchmarks, d1-3B averaged 82.9, surpassing Decider 4B's 81.1. The d1-omni-600M model averaged 78.4, achieving top scores on Civil Comments (95.8) and PAWS-X (79.5). In 11 image benchmarks, d1-3B averaged 74.1, a marginal improvement over its base model's 73.9. The company notes that audio decision benchmarks remain an active area of research.
NVIDIA Hardware Latency
Performance measurements on NVIDIA hardware indicate high efficiency. A single d1-3B question takes 8 milliseconds on an RTX 4090 when using specific compilation modes, increasing to 16 milliseconds without. An AMD MI325X processes a question in 9 milliseconds. On Jetson edge devices, an AGX Thor takes 16 milliseconds for one question, expanding to 20 milliseconds for three questions over a single state. An AGX Orin completes the task in 26 milliseconds, and an Orin Nano in 50 milliseconds. NVIDIA's Jetson AI Lab also offers dedicated guides for d1-3B.
Local Deployment Guidelines
For local execution, d1-3B requires transformers version 5.14 or newer, while d1-omni-600M necessitates transformers version 5.15 or newer, with `trust_remote_code=True` enabled. Both models offer `system_one(state, questions)` for individual requests and `system_one_batch` for grouped processing. For optimal accuracy of d1-omni-600M on a GPU, utilizing `float16` is recommended as `bfloat16` was observed to alter some top answers. Users can explore practical applications through ten camera-driven d1-3B demonstrations available in the Open d1 Arcade space.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
