Skip to content

LFM3 VL 2B VRAM requirements

How much VRAM LFM3 VL 2B needs
4-bit weights: 1.1 GBWith 8K context: 1.9 GBWith 32K context: 2.2 GB8-bit, 8K context: 2.8 GB
Parameters
1.9B
Longest context
125K
Released
Oct 2026
License
Lfm1.0
Strongest at
Text and images

Its licence (Lfm1.0) is not a standard open-source one: check the terms before commercial use.

How much VRAM LFM3 VL 2B needs

Memory LFM3 VL 2B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 125K
16-bit3.6 GB4.4 GB4.7 GB5.8 GB
8-bit1.9 GB2.8 GB3.0 GB4.1 GB
6-bit1.5 GB2.3 GB2.6 GB3.7 GB
5-bit1.3 GB2.1 GB2.4 GB3.5 GB
4-bit1.1 GB1.9 GB2.2 GB3.3 GB
3-bit894 MB1.7 GB2.0 GB3.1 GB
2-bit686 MB1.5 GB1.8 GB2.9 GB
1.6-bit377 MB1.2 GB1.5 GB2.6 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run LFM3 VL 2B

The best way LFM3 VL 2B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit125K88 tok/s0.2 sPDF: LFM3 VL 2B on RTX 4060 report
RTX 50608-bit125K139 tok/s0.2 sPDF: LFM3 VL 2B on RTX 5060 report
RTX 30708-bit125K139 tok/s0.2 sPDF: LFM3 VL 2B on RTX 3070 report
RTX 3080 10 GB8-bit125K220 tok/s0.1 sPDF: LFM3 VL 2B on RTX 3080 10 GB report
RTX 3060 12 GB8-bit125K114 tok/s0.3 sPDF: LFM3 VL 2B on RTX 3060 12 GB report
Arc B5808-bit125K96 tok/s0.2 sPDF: LFM3 VL 2B on Arc B580 report
RTX 40708-bit125K154 tok/s0.1 sPDF: LFM3 VL 2B on RTX 4070 report
RTX 4070 Super8-bit125K154 tok/s0.1 sPDF: LFM3 VL 2B on RTX 4070 Super report
RTX 50708-bit125K198 tok/s0.1 sPDF: LFM3 VL 2B on RTX 5070 report
RTX 4060 Ti 16 GB8-bit125K92 tok/s0.2 sPDF: LFM3 VL 2B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit125K139 tok/s0.2 sPDF: LFM3 VL 2B on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit125K115 tok/s0.2 sPDF: LFM3 VL 2B on Arc A770 16 GB report
Radeon RX 9070 XT8-bit125K154 tok/s0.1 sPDF: LFM3 VL 2B on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit125K198 tok/s0.1 sPDF: LFM3 VL 2B on RTX 4070 Ti Super report
RTX 4080 Super8-bit125K214 tok/s0.1 sPDF: LFM3 VL 2B on RTX 4080 Super report
RTX 5070 Ti8-bit125K252 tok/s0.1 sPDF: LFM3 VL 2B on RTX 5070 Ti report
RTX 50808-bit125K266 tok/s0.1 sPDF: LFM3 VL 2B on RTX 5080 report
Radeon RX 7900 XTX8-bit125K211 tok/s0.2 sPDF: LFM3 VL 2B on Radeon RX 7900 XTX report
RTX 30908-bit125K261 tok/s0.1 sPDF: LFM3 VL 2B on RTX 3090 report
RTX 40908-bit125K277 tok/s0.1 sPDF: LFM3 VL 2B on RTX 4090 report
Radeon AI PRO R97008-bit125K154 tok/s0.1 sPDF: LFM3 VL 2B on Radeon AI PRO R9700 report
RTX 50908-bit125K426 tok/s0.1 sPDF: LFM3 VL 2B on RTX 5090 report
2× RTX 30908-bit125K261 tok/s0.1 sPDF: LFM3 VL 2B on 2× RTX 3090 report
2× RTX 40908-bit125K277 tok/s0.1 sPDF: LFM3 VL 2B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit125K346 tok/s0.1 sPDF: LFM3 VL 2B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit125K266 tok/s0.1 sPDF: LFM3 VL 2B on RTX 6000 Ada report
L40S8-bit125K244 tok/s0.1 sPDF: LFM3 VL 2B on L40S report
2× RTX 50908-bit125K426 tok/s0.1 sPDF: LFM3 VL 2B on 2× RTX 5090 report
4× RTX 30908-bit125K261 tok/s0.1 sPDF: LFM3 VL 2B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit125K426 tok/s0.1 sPDF: LFM3 VL 2B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit125K44 tok/s2.4 sPDF: LFM3 VL 2B on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit125K44 tok/s2.4 sPDF: LFM3 VL 2B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit125K44 tok/s2.4 sPDF: LFM3 VL 2B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit125K91 tok/s1.2 sPDF: LFM3 VL 2B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit125K91 tok/s1.2 sPDF: LFM3 VL 2B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit125K91 tok/s1.2 sPDF: LFM3 VL 2B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit125K145 tok/s0.6 sPDF: LFM3 VL 2B on M4 Max 128 GB report
M5 Max 128 GB8-bit125K158 tok/s0.2 sPDF: LFM3 VL 2B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit125K174 tok/s0.3 sPDF: LFM3 VL 2B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit125K174 tok/s0.3 sPDF: LFM3 VL 2B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit125K174 tok/s0.3 sPDF: LFM3 VL 2B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit125K221 tok/s0.1 sPDF: LFM3 VL 2B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit125K69 tok/s0.3 sPDF: LFM3 VL 2B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit125K69 tok/s0.3 sPDF: LFM3 VL 2B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit125K96 tok/s0.1 sPDF: LFM3 VL 2B on RTX Spark report
DGX Spark8-bit125K88 tok/s0.1 sPDF: LFM3 VL 2B on DGX Spark report
Dual DGX Spark8-bit125K88 tok/s0.1 sPDF: LFM3 VL 2B on Dual DGX Spark report
DGX Station GB3008-bit125K885 tok/s0.1 sPDF: LFM3 VL 2B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit125K426 tok/s0.1 sPDF: LFM3 VL 2B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit125K426 tok/s0.1 sPDF: LFM3 VL 2B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit125K426 tok/s0.1 sPDF: LFM3 VL 2B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit125K466 tok/s0.1 sPDF: LFM3 VL 2B on A100 80 GB SXM report
H100 SXM8-bit125K629 tok/s0.1 sPDF: LFM3 VL 2B on H100 SXM report
4× H200 NVL8-bit125K754 tok/s0.1 sPDF: LFM3 VL 2B on 4× H200 NVL report
8× H200 NVL8-bit125K754 tok/s0.1 sPDF: LFM3 VL 2B on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

llama.cpp
llama-server -hf LiquidAI/LFM3-VL-2B-GGUF
vLLM
vllm serve LiquidAI/LFM3-VL-2B
MLX (Mac)
mlx_lm.chat --model mlx-community/LFM3-VL-2B-OptiQ-4bit

Questions

How much VRAM does LFM3 VL 2B need?

LFM3 VL 2B needs 1.1 GB for its weights at 4-bit, 1.9 GB with an 8K context and 2.2 GB with 32K. At 8-bit, 2.8 GB.

What is the minimum VRAM for LFM3 VL 2B?

About 1.7 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the RTX 4060.

How big is LFM3 VL 2B?

1.9B parameters: 3.6 GB of weights at 16-bit, 1.1 GB at 4-bit.

Can LFM3 VL 2B run on an RTX 4090?

Yes. At 8-bit it holds 125K of context and writes about 277 tokens a second.

More models