Skip to content

Back to the catalog

Liquid AI (@liquidai) on X
x.comthread

Liquid AI (@liquidai) on X

Liquid AI announces Open d1, two open-weight multimodal decision models. d1-3B (built from LFM2.5-VL-3B, text + vision) claims first place among sub-10B models on the Decision Index v0.2.1, while the experimental d1-omni-600M (LFM2.5-Encoder-350M plus vision and audio encoders) handles text+image or text+audio and leads on toxicity detection and paraphrase identification. Single-request d1-3B latency is 8 ms on an RTX 4090, 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin 64 GB and 50 ms on Jetson Orin Nano, with day-one llama.cpp support; demos include live-camera apps and Isaac Sim navigation. Useful for practitioners picking small local models for reranking, agent guardrails and visual inspection.

TL;DR Liquid AI releases open-weight d1-3B and d1-omni-600M multimodal decision models with edge latency numbers on RTX/Jetson and llama.cpp support.
Open original ↗
Document

What it claims

Liquid AI's Open d1 release consists of two open-weight decision models, distributed as downloadable weights for local fine-tuning and deployment:

  • d1-3B takes text and images. It is built from LFM2.5-VL-3B, makes a decision in one pass, and is pitched for reranking, agent guardrails and visual inspection.
  • d1-omni-600M is labelled experimental. It takes text plus image, or text plus audio, and is pitched for voice-command routing, among other uses that the thread truncates. Liquid AI recommends it "when footprint matters" and recommends d1-3B for "high-quality text and vision decisions".

Evidence

  • Decision Index ranking. d1-3B ranks first among models under 10B on Decision Index v0.2.1, which the thread calls "a benchmark for structured decision-making". The thread gives no scores and names no competing models.
  • d1-omni-600M results. It "leads our text benchmark comparison" in toxicity detection and paraphrase identification. This is Liquid AI's own comparison, run on text tasks only, so it says nothing about the image or audio paths.
  • Demos. There are 10 live-camera demos for d1-3B, ranging from gesture-controlled games to content moderation, each running one forward pass per frame. An Isaac Sim navigation demo, built with NVIDIA Robotics, serves the model on a Jetson.

How to use it

  • Weights for d1-3B are at huggingface.co/LiquidAI/d1-3b. The thread cuts off before the d1-omni-600M link.
  • The release blog post is at liquid.ai/blog/d1-open.
  • llama.cpp support is available from day one, and the thread lists NVIDIA DGX, RTX and Jetson as supported targets.

Limits

  • Truncated source. This is an X thread, and several posts are cut off at "Show more", including the d1-omni-600M use cases, the d1-omni weights link and the Isaac Sim details. The blog post is likely the fuller source.
  • No licence. The thread says "open-weight" but states no licence. Check the Hugging Face model card before adopting.
  • Latency figures measure one request at a time. They are not throughput under batching or concurrency, and the thread does not give quantization, precision or input size.
  • d1-omni-600M is flagged as experimental, and its stated advantages come from a vendor-run text-only comparison.

Organisations named here

  • Liquid AI
  • NVIDIA
  • NVIDIA Robotics
  • Decision Index v0.2.1
  • Hugging Face

Technologies named here

  • llama.cpp
  • d1-3B
  • Isaac Sim
  • d1-omni-600M
  • NVIDIA Jetson
  • LFM2.5-VL-3B
  • LFM2.5-Encoder-350M
  • NVIDIA DGX
  • NVIDIA RTX
  • Hugging Face
  • Vision language model

Techniques named here

  • Single forward pass per frame decision-making

Related in models