Skip to content

LFM2.5 8B A1B VRAM requirements

How much VRAM LFM2.5 8B A1B needs
4-bit weights: 4.8 GBWith 8K context: 5.6 GBWith 32K context: 5.9 GB8-bit, 8K context: 9.2 GB
Parameters
8.5B
Active per token
1B
Longest context
125K
Released
May 2026
License
Lfm1.0
Strongest at
Not rated

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

How much VRAM LFM2.5 8B A1B needs

Memory LFM2.5 8B A1B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 125K
16-bit15.8 GB16.6 GB16.9 GB18.0 GB
8-bit8.4 GB9.2 GB9.5 GB10.6 GB
6-bit6.5 GB7.3 GB7.6 GB8.7 GB
5-bit5.6 GB6.5 GB6.7 GB7.8 GB
4-bit4.8 GB5.6 GB5.9 GB7.0 GB
3-bit3.9 GB4.7 GB5.0 GB6.1 GB
2-bit3.0 GB3.8 GB4.1 GB5.2 GB
1.6-bit1.6 GB2.5 GB2.8 GB3.8 GB

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

GPUs and Macs that run LFM2.5 8B A1B

The best way LFM2.5 8B A1B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40606-bit32K174 tok/s0.4 sPDF: LFM2.5 8B A1B on RTX 4060 report
RTX 50606-bit32K247 tok/s0.3 sPDF: LFM2.5 8B A1B on RTX 5060 report
RTX 30706-bit32K247 tok/s0.3 sPDF: LFM2.5 8B A1B on RTX 3070 report
RTX 3080 10 GB8-bit28K291 tok/s0.2 sPDF: LFM2.5 8B A1B on RTX 3080 10 GB report
RTX 3060 12 GB8-bit125K177 tok/s0.5 sPDF: LFM2.5 8B A1B on RTX 3060 12 GB report
Arc B5808-bit125K142 tok/s0.3 sPDF: LFM2.5 8B A1B on Arc B580 report
RTX 40708-bit125K225 tok/s0.2 sPDF: LFM2.5 8B A1B on RTX 4070 report
RTX 4070 Super8-bit125K225 tok/s0.2 sPDF: LFM2.5 8B A1B on RTX 4070 Super report
RTX 50708-bit125K270 tok/s0.2 sPDF: LFM2.5 8B A1B on RTX 5070 report
RTX 4060 Ti 16 GB8-bit125K149 tok/s0.3 sPDF: LFM2.5 8B A1B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit125K207 tok/s0.3 sPDF: LFM2.5 8B A1B on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit125K161 tok/s0.3 sPDF: LFM2.5 8B A1B on Arc A770 16 GB report
Radeon RX 9070 XT8-bit125K203 tok/s0.2 sPDF: LFM2.5 8B A1B on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit125K270 tok/s0.1 sPDF: LFM2.5 8B A1B on RTX 4070 Ti Super report
RTX 4080 Super8-bit125K285 tok/s0.1 sPDF: LFM2.5 8B A1B on RTX 4080 Super report
RTX 5070 Ti8-bit125K319 tok/s0.1 sPDF: LFM2.5 8B A1B on RTX 5070 Ti report
RTX 50808-bit125K331 tok/s0.1 sPDF: LFM2.5 8B A1B on RTX 5080 report
Radeon RX 7900 XTX8-bit125K250 tok/s0.3 sPDF: LFM2.5 8B A1B on Radeon RX 7900 XTX report
RTX 30908-bit125K327 tok/s0.2 sPDF: LFM2.5 8B A1B on RTX 3090 report
RTX 40908-bit125K339 tok/s0.1 sPDF: LFM2.5 8B A1B on RTX 4090 report
Radeon AI PRO R97008-bit125K203 tok/s0.2 sPDF: LFM2.5 8B A1B on Radeon AI PRO R9700 report
RTX 50908-bit125K437 tok/s0.1 sPDF: LFM2.5 8B A1B on RTX 5090 report
2× RTX 30908-bit125K327 tok/s0.1 sPDF: LFM2.5 8B A1B on 2× RTX 3090 report
2× RTX 40908-bit125K339 tok/s0.1 sPDF: LFM2.5 8B A1B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit125K389 tok/s0.1 sPDF: LFM2.5 8B A1B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit125K331 tok/s0.1 sPDF: LFM2.5 8B A1B on RTX 6000 Ada report
L40S8-bit125K313 tok/s0.1 sPDF: LFM2.5 8B A1B on L40S report
2× RTX 50908-bit125K437 tok/s0.1 sPDF: LFM2.5 8B A1B on 2× RTX 5090 report
4× RTX 30908-bit125K327 tok/s0.1 sPDF: LFM2.5 8B A1B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit125K437 tok/s0.1 sPDF: LFM2.5 8B A1B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit96K71 tok/s3.8 sPDF: LFM2.5 8B A1B on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit125K71 tok/s3.8 sPDF: LFM2.5 8B A1B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit125K71 tok/s3.8 sPDF: LFM2.5 8B A1B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit125K123 tok/s1.9 sPDF: LFM2.5 8B A1B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit125K123 tok/s1.9 sPDF: LFM2.5 8B A1B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit125K123 tok/s1.9 sPDF: LFM2.5 8B A1B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit125K168 tok/s1.0 sPDF: LFM2.5 8B A1B on M4 Max 128 GB report
M5 Max 128 GB8-bit125K176 tok/s0.3 sPDF: LFM2.5 8B A1B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit125K187 tok/s0.5 sPDF: LFM2.5 8B A1B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit125K187 tok/s0.5 sPDF: LFM2.5 8B A1B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit125K187 tok/s0.5 sPDF: LFM2.5 8B A1B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit125K212 tok/s0.2 sPDF: LFM2.5 8B A1B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit125K110 tok/s0.6 sPDF: LFM2.5 8B A1B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit125K110 tok/s0.6 sPDF: LFM2.5 8B A1B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit125K154 tok/s0.2 sPDF: LFM2.5 8B A1B on RTX Spark report
DGX Spark8-bit125K143 tok/s0.2 sPDF: LFM2.5 8B A1B on DGX Spark report
Dual DGX Spark8-bit125K143 tok/s0.1 sPDF: LFM2.5 8B A1B on Dual DGX Spark report
DGX Station GB3008-bit125K605 tok/s0.1 sPDF: LFM2.5 8B A1B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit125K437 tok/s0.1 sPDF: LFM2.5 8B A1B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit125K437 tok/s0.1 sPDF: LFM2.5 8B A1B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit125K437 tok/s0.1 sPDF: LFM2.5 8B A1B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit125K458 tok/s0.1 sPDF: LFM2.5 8B A1B on A100 80 GB SXM report
H100 SXM8-bit125K528 tok/s0.1 sPDF: LFM2.5 8B A1B on H100 SXM report
4× H200 NVL8-bit125K569 tok/s0.1 sPDF: LFM2.5 8B A1B on 4× H200 NVL report
8× H200 NVL8-bit125K569 tok/s0.1 sPDF: LFM2.5 8B A1B 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/LFM2.5-8B-A1B-GGUF
vLLM
vllm serve LiquidAI/LFM2.5-8B-A1B
MLX (Mac)
mlx_lm.chat --model mlx-community/LFM2.5-8B-A1B-MLX-4bit

Questions

How much VRAM does LFM2.5 8B A1B need?

LFM2.5 8B A1B needs 4.8 GB for its weights at 4-bit, 5.6 GB with an 8K context and 5.9 GB with 32K. At 8-bit, 9.2 GB.

What is the minimum VRAM for LFM2.5 8B A1B?

About 4.7 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 LFM2.5 8B A1B?

8.5B parameters (1B active per token): 15.8 GB of weights at 16-bit, 4.8 GB at 4-bit.

Can LFM2.5 8B A1B run on an RTX 4090?

Yes. At 8-bit it holds 125K of context and writes about 339 tokens a second.

More models