Liquid AI Opens d1-3B: Edge Decision Model Answers in 8ms on an RTX 4090, Outscores a 35B Rival
Liquid AI released two open-weight models on October 7: d1-3B and d1-omni-600M, the first entries in its d1 decision model family. Both are available on Hugging Face under open weights.
Decision models are specialized for classification, routing, and real-time agentic decisions rather than open-ended generation. Where a frontier model handles a task end-to-end, a decision model handles the microsecond question in front of it: which tool to call, which queue to route to, whether an input meets a threshold. The d1 series uses Liquid AI’s non-transformer architecture, optimized for minimum latency rather than throughput at scale.
d1-3B: beating a model 12x its size
d1-3B scores 48.57 on Decision Index v0.2.1 (public split), ahead of Decider 35B-A3B and on par with Decider 35B. On seven text benchmarks covering reading comprehension, toxicity detection, intent classification, medical QA, and cross-lingual understanding, d1-3B posts a mean of 82.9 — above Decider 4B at 81.1.
Latency is the headline:
| Device | Latency (single request) |
|---|---|
| NVIDIA RTX 4090 | 8ms |
| Jetson AGX Thor | 16ms |
| Jetson AGX Orin 64GB | 26ms |
| Jetson Orin Nano | 50ms |
| Apple M5 Pro | measured separately |
At 50ms on a Jetson Orin Nano, d1-3B runs fast enough for real-time decisions on the smallest commercially available Jetson hardware. Liquid AI ran these benchmarks in collaboration with NVIDIA.
d1-omni-600M
d1-omni-600M is an early experimental multimodal checkpoint handling text, image, and audio at 600 million parameters. It scores 15.95 on the Decision Index and 78.4 on the text benchmark mean, outperforming Decider 2B at one-quarter the parameters. Liquid AI flags it as under active development.
On toxicity detection (Civil Comments), d1-omni-600M scores 95.8 — the highest in its benchmark table. On paraphrase identification (PAWS-X), 79.5, also the table’s highest.
Why it matters for agentic systems
Most agent architectures call a large model for routing decisions that do not require frontier-level reasoning. Each network round-trip adds 50ms to several hundred milliseconds of latency. A 3B model that routes, classifies, or triggers at 8ms on a GPU or 50ms on an edge board collapses that latency to background noise, enabling agentic systems with hard real-time requirements.
The MIT license removes deployment friction. Liquid AI said GGUF, MLX, and similar conversion formats are forthcoming for the models.