// Liquid LFM · Vision-language
LFM2.5-VL-3B: VRAM & local deployment
3.12B parameters. About 2.7 GB at 4-bit and 7.5 GB at BF16 with one 8K-token request.
released 2026-08-11licence: Custom licence – read itmemory figures are estimates
01Specs
| Publisher | Liquid AI · LiquidAI |
|---|---|
| Family / series | Liquid LFM · LFM2.5-VL |
| Type | Vision-language (image-text-to-text) |
| Parameters | 3.12B Source: counted from official safetensors weights |
| Context length | 32K tokens (32,768)max_position_embeddings in config.json |
| Architecture | 30 layers (8 full-attention) · GQA+HYBRID |
| Released weights | BF16 · 6.3 GBdownload |
| Licence | lfm1.0 · Custom licence – read itThe publisher uses its own licence; we have not classified it. Read the licence text. |
| Released | 2026-08-11 (Hugging Face repo created) |
02How much VRAM it needs (estimate)
Total = weights + 10% + 0.5GB overhead + KV cache for one request (FP16 KV). More concurrent users multiply the KV column.
| Precision | Weights | Total @ 8K | Total @ 32K | Smallest GPU | Mac |
|---|---|---|---|---|---|
| BF16 / FP16 (original) | 6.3 GB | 7.5 GB | 7.9 GB | 12GB GPU | Mac 16GB |
| FP8 / INT8 / Q8_0 | 3.3 GB | 4.3 GB | 4.7 GB | 8GB GPU | Mac 16GB |
| 4-bit (Q4_K_M / AWQ / GPTQ-Int4) | 1.9 GB | 2.7 GB | 3.1 GB | 8GB GPU | Mac 16GB |
KV cache per request: 8K 0.1 GB · 32K 0.5 GB
Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →
03How to run it
Only runtimes the publisher or the runtime’s official library names for this model.
- vLLM — mentioned in the official model card
- SGLang — mentioned in the official model card
- llama.cpp — mentioned in the official model card
- MLX — mentioned in the official model card
- Transformers — mentioned in the official model card
04Official repositories & quantised builds
8 official repo(s) from LiquidAI.
- LFM2.5-VL-3B-GGUF GGUF
- LFM2.5-VL-3B main
- LFM2.5-VL-3B-MLX-4bit MLX
- LFM2.5-VL-3B-ONNX ONNX
- LFM2.5-VL-3B-MLX-8bit MLX
- LFM2.5-VL-3B-MLX-bf16 MLX/BF16
- LFM2.5-VL-3B-MLX-6bit MLX
- LFM2.5-VL-3B-MLX-5bit MLX
05Other Liquid LFM models
All 26 →| Model | Params | 4-bit total | Released |
|---|---|---|---|
| LFM2.5-VL-450M · same series | 449M | 0.9 GB | 2026-04-08 |
| LFM2.5-VL-1.6B · same series | 1.6B | 1.7 GB | 2026-01-05 |
| LFM2.5-24B-A2B | 24B / 2B | 16.7 GB | 2026-10-12 |
| LFM2.5-2.6B | 2.7B | 2.4 GB | 2026-07-28 |
| LFM2.5-Encoder-350M-Diffusion | 354M | 0.8 GB | 2026-07-10 |
| LFM2.5-230M | 230M | 0.8 GB | 2026-06-24 |
| LFM2.5-8B-A1B | 8.47B / 1B | 6.2 GB | 2026-05-28 |
| LFM2.5-1.2B-JP-202606 | 1.17B | 1.4 GB | 2026-05-26 |
| LFM2.5-ColBERT-350M | 353M | 0.7 GB | 2026-05-20 |
| LFM2.5-Embedding-350M | 354M | 0.7 GB | 2026-05-05 |
06Sources
Data checked 2026-09-24. Parameters, context and licence are copied from the official repository; memory figures are our estimates.