d1-3B — VRAM & Python setup
Written by Jakub Rusinowski · Last updated
Liquid AI's 3B decision model on LFM2.5-VL-3B, for text and images, built for the edge: 8 ms a question on an RTX 4090 and 50 ms on a Jetson Orin Nano (vendor figures). A Python library, not a server.
d1-3B needs about 3 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
Call d1-3B
A Python library: you load it inside your own program. It has no HTTP server and does not run in Ollama.
- Question types
- choice noul (yes / no) score
- Inputs
- text image
- Serve with
Transformers (trust_remote_code)
pip install "transformers>=5.14" torch torchvision pillowNot in the request builder: d1 is a Python library with no HTTP endpoint.
Related guides: Confidence, thresholds and calibration · Decision models on images
Other decision models: Nimble 9B · Tev1 4B · Tev1 0.8B · Winnow 12B · Winnow E4B · Decider 2B · Decider 4B · Decider 35B-A3B (NVFP4) · JevK5 4B · Intern-Decision 4B · AutoJev 27B · Laya · Clef 27B · Clef-Flash 9B · Rune 26B-A4B v3 · OpenDecider-nano · OpenDecider-small · OpenDecider-small MLX 4-bit · Bosun v3.1 0.6B · Bosun v3.1 1.7B · d1-omni-600M · THX-01 322M
Hardware fit
Weights plus overhead plus the KV cache at a 8,192-token prompt, on NVIDIA RTX 4090 (24 GB). A publisher build is sized from its file; the other rows are modelled at a standard quant. Decision requests are short, so no long-context figure is shown. Image tokens are not modelled: each image adds to the prompt, so a request with images needs more memory than these rows.
| Quant | Memory | VRAM | Fit |
|---|---|---|---|
| BF16 Publisher build · 6.25 GB file | 7.9 GB | Fits | |
| Q4_K_M 4.83 bpw · modelled quant | 3.5 GB | Fits | |
| Q6_K 6.56 bpw · modelled quant | 4.2 GB | Fits | |
| Q8_0 8.50 bpw · modelled quant | 5 GB | Fits |
Published file size: BF16 6.25 GB. A download size from the model publisher — not a VRAM requirement.
How it was scored
Each table is one suite. Numbers in different tables were measured on different questions and are not comparable, with each other or with the Decision Index.
Decision Index 0.2.1 (snapshot 2026-09-28)
Decision Index 0.2.1Community Decision Index (not by the model's authors)d1 is not on the 28 September 2026 snapshot. Liquid AI scored it with the official scorer itself; that run is shown in its own table below and is not a leaderboard result. Community-maintained; not affiliated with the model authors.
Source ↗| What was scored | Score | Source |
|---|---|---|
| Balanced skill · 0.2.1 | 48.57 | Source ↗ |
Not a leaderboard submission, and not the community snapshot this site uses: Liquid AI ran the scorer itself.
| What was scored | Score | Source |
|---|---|---|
| Mean of 7 (SQuAD 2.0, Civil Comments, MASSIVE, PubMedQA, BoolQ, XNLI, PAWS-X) | 82.9% | Source ↗ |
| Mean of 8 (the 7 above plus HelpSteer2) | 77.1% | Source ↗ |
| What was scored | Score | Source |
|---|---|---|
| Mean, at most 1,000 rows each | 74.1% | Source ↗ |
Published latency
| Device | Time | Condition | Who measured it | Source |
|---|---|---|---|---|
| NVIDIA RTX 4090 | 8 ms | 1 question, warm, bf16, compiled as CUDA graphs (16 ms without) | Authors' own internal benchmark — not independently reproduced | Source ↗ |
| AMD MI325X | 9 ms | 1 question, warm, bf16 | Authors' own internal benchmark — not independently reproduced | Source ↗ |
| NVIDIA Jetson AGX Thor | 16 ms | 1 question, warm | Authors' own internal benchmark — not independently reproduced | Source ↗ |
| NVIDIA Jetson AGX Orin 64 GB | 26 ms | 1 question, warm | Authors' own internal benchmark — not independently reproduced | Source ↗ |
| NVIDIA Jetson Orin Nano | 50 ms | 1 question, warm | Authors' own internal benchmark — not independently reproduced | Source ↗ |
| Apple M5 Pro (mps) | 30 ms | 1 question, warm | Authors' own internal benchmark — not independently reproduced | Source ↗ |
As published, on the device and under the condition shown. A decision model answers in one forward pass, so this is time per decision, not tokens per second.
Decision models are not ranked on chat, creative or coding scores.
Specifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- Parameters
- 3.12B
- Context window
- 32K (32,768)
- Architecture
- Fine-tune of LiquidAI/LFM2.5-VL-3B
- Provider
- Liquid AI
- Licence
- LFM Open License v1.0
- Availability
- Weights downloadable
- Practical deployment
- Edge device (phone, SBC)
- Specified at
- Q4_K_M
- System RAM
- 8 GB
- Record updated
- 2026-10-13
- Runs with
- Transformers: supported
Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.
Other d1 sizes
d1-3B — frequently asked questions
What is d1-3B?
d1-3B is a decision model: you send it a state and typed questions (choice, yes/no/unknown, or a score) and it returns one answer per question with a probability for every option. It is not a chat model.
How do I run d1-3B locally?
A Python library: you load it inside your own program. It has no HTTP server and does not run in Ollama. Install: `pip install "transformers>=5.14" torch torchvision pillow`.
How much memory does d1-3B need?
About 3 GB for the weights plus overhead at Q4_K_M, before the prompt's KV cache. Decision prompts are short, so the cache stays small.
How accurate is d1-3B?
It has no community Decision Index score. Every score on this page comes from its authors, on the suites named beside each table; none is an independent result.