Model reference · open weights
LFM3-VL is an open-weight language model from LiquidAI. LFM3-VL-2B (BF16) weighs 3.8 GB; the smallest configuration that runs it is RTX 3060 12 GB.
For self-hosting, compare this build's requirements below with the GPU you plan to rent.
Summary of the LiquidAI/LFM3-VL-2B-GGUF model card, 2026-10-29. The estimate below is for another build of the family.
What it is
| Released by | LiquidAI |
|---|---|
| Released | 2026-10-19 |
| VRAM | 3.8 GB for the weights |
What it runs on
| Card | Requests at once | Context max | Memory | |
|---|---|---|---|---|
| 8K each | 32K each | |||
| RTX 3060 12 GB | 16 | 4 | all 125K | 11.6 GB |
| RTX 4060 Ti 16 GB | 26 | 7 | all 125K | 15.4 GB |
| RTX 3090 24 GB | 48 | 13 | all 125K | 23.4 GB |
| RTX 4090 24 GB | 48 | 13 | all 125K | 23.4 GB |
| RTX 5090 32 GB | 69 | 19 | all 125K | 31.0 GB |
| L40S 48 GB | 105 | 28 | all 125K | 44.0 GB |
| A100 80 GB | 199 | 54 | all 125K | 78.2 GB |
| H100 80 GB | 170 | 47 | all 125K | 78.1 GB |
| RTX PRO 6000 Blackwell 96 GB | 209 | 58 | all 125K | 93.8 GB |
| DGX Spark (GB10) 128 GB unified | 242 | 67 | all 125K | 107 GB |
| H200 141 GB | 318 | 88 | all 125K | 138 GB |
| B200 180 GB | 412 | 103 | all 125K | 176 GB |
| Requests at once | 8K tokens each | 32K tokens each |
|---|---|---|
| 1 | 6.1 GB | 7.1 GB |
| 5 | 7.6 GB | 12.4 GB |
| 8 | 8.7 GB | 16.3 GB |
| 16 | 11.6 GB | 26.9 GB |
| 32 | 17.4 GB | 48.0 GB |
| 64 | 29.0 GB | 90.2 GB |
One card, with vLLM's small-card settings.
Builds
| Build | Params | Precision | Weights | Smallest setup |
|---|---|---|---|---|
| LFM3-VL-2B-GGUF ↗ 7 builds: Q4_0 1.0 GB … F16 3.5 GB |
GGUF | 1.1 GB | RTX 3060 12 GB | |
| LFM3-VL-2B (above) ↗ | 1.9B | BF16 | 3.8 GB | RTX 3060 12 GB |
From the model card
src="https://huggingface.co/LiquidAI/LFM3-1.7B/resolve/main/assets/liquid-ai-banner.webp" alt="Liquid AI" style="width: 100%; max-width: 100%; height: auto; display: block; margin: 0;" />
GGUF version of LFM3-VL-2B for local inference on CPUs and GPUs.
Find architecture and usage details in the original model card.
Recommended sampling: temperature 0.1, top_k 50, top_p 1.0, min_p 0.0, repeat penalty 1.1.
ollama run lfm3-vl, for macOS: image input. Requires Ollama version v0.40.12 or newer.The Q4_0 build in this repository is the Quantization-Aware Distillation (QAD) checkpoint,
LFM3-VL-2B-QAD-Q4_0.gguf.
It was trained to recover the accuracy lost to Q4_0 quantization.
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.
How it works