d1-omni 600M VRAM requirements

Weight memory by runtime

GiB means 1,073,741,824 bytes. FP16 storage is parameters × 2 bytes. Published FP32 file bytes include safetensors headers. The model file is downloaded in FP32 even when GPU tensors are loaded as FP16.

24 and 32 GiB GPUs comfortably exceed the weight footprint. An 8 GiB GPU also has room for the FP16 weights, but long inputs can exceed its capacity: our 7,075-token fixture reached 7.19 GiB of peak allocated CUDA tensors, with nvidia-smi reporting 8.28 GiB after the call. Input length, image tiles, audio duration and batching determine the remaining memory.

ModelVRAM measured this on October 9, 2026, using one RTX PRO 6000 Blackwell Server Edition, PyTorch 2.8.0, CUDA 12.8 and Transformers 5.17.0. FP16 text requests of 45–46 tokens used 1.12 GiB peak allocated memory and took 7.18–7.31 ms; 7,075 tokens took 66.00 ms. Times are medians of three calls after one warmup. CUDA allocated memory excludes driver overhead and unused allocator reservations. Inputs, timings, memory and checkpoint SHA256; measurement script.

What is included in the 600M model?

The publisher rounds the full checkpoint to 587M parameters: a 381M shared trunk and decision head, a 94M SigLIP2 vision encoder and a 112M FastConformer audio encoder. This page counts all three, matching the single model.safetensors file. The 350M base encoder is not the complete deployment footprint.

Its overall input limit is 16,384 positions. With images, the loader restricts state and question text to 896 tokens. Audio requests accept one 16 kHz mono clip, capped at 30 seconds; a request cannot combine images and audio. More than one image may create extra tiles and therefore extra working memory.

These are single-pass decision workloads. The normal chat calculator’s growing generation KV cache and output tokens per second do not describe system_one. LiquidAI does not publish inference speed numbers for this early research release.

Run the official Python API

pip install "transformers>=5.15" torch torchvision pillow soundfile

import torch
from transformers import AutoModel

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if device == "cuda" else torch.float32
model = AutoModel.from_pretrained(
    "LiquidAI/d1-omni-600M", trust_remote_code=True, dtype=dtype
).to(device)
questions = {
    "team": {
        "type": "choice",
        "instructions": "Choose a team for this ticket.",
        "criteria": {"billing": "Payment issues", "technical": "Software faults"},
    }
}
print(model.system_one("The application crashes at startup.", questions))

The official loader requires trust_remote_code=True. Review the repository’s Python implementation before loading it. ModelVRAM exercised this API on CUDA with FP16 and on CPU with FP32, including text, image, audio and batch requests. Eight handwritten text tickets were classified correctly; four synthetic color images produced the same green answer in both FP16 and FP32. These are smoke fixtures, not a representative accuracy benchmark. The one-second silence fixture only checks the audio API. All smoke outputs.

License and sources

The repository uses Liquid AI LFM License 1.0. Check its commercial-use conditions for your deployment; the radar’s earlier “other” label is not the license name.

Read on October 9, 2026: LiquidAI’s model card, config.json and the FP32 checkpoint. Compare with Perplexity’s 27B Decider and Laya’s text decision models. How estimates and measurements are reported.