Skip to content

Lift VRAM requirements

How much VRAM Lift needs
4-bit weights: 5.5 GBWith 8K context: 6.8 GBWith 32K context: 7.5 GB8-bit, 8K context: 10.9 GB
Parameters
9.7B
Longest context
256K
Released
Jun 2026
License
OpenRAIL
Strongest at
Text and images

How much VRAM Lift needs

Memory Lift needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit18.0 GB19.3 GB20.1 GB23.1 GB
8-bit9.6 GB10.9 GB11.6 GB14.6 GB
6-bit7.4 GB8.7 GB9.5 GB12.5 GB
5-bit6.4 GB7.7 GB8.5 GB11.5 GB
4-bit5.5 GB6.8 GB7.5 GB10.5 GB
3-bit4.4 GB5.7 GB6.5 GB9.5 GB
2-bit3.4 GB4.7 GB5.5 GB8.5 GB
1.6-bit1.9 GB3.2 GB4.0 GB7.0 GB

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

GPUs and Macs that run Lift

The best way Lift runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40604-bit32K35 tok/s1.1 sPDF: Lift on RTX 4060 report
RTX 50604-bit32K56 tok/s0.9 sPDF: Lift on RTX 5060 report
RTX 30704-bit32K56 tok/s0.8 sPDF: Lift on RTX 3070 report
RTX 3080 10 GB6-bit32K70 tok/s0.6 sPDF: Lift on RTX 3080 10 GB report
RTX 3060 12 GB8-bit20K27 tok/s1.3 sPDF: Lift on RTX 3060 12 GB report
Arc B5808-bit20K24 tok/s0.9 sPDF: Lift on Arc B580 report
RTX 40708-bit20K37 tok/s0.6 sPDF: Lift on RTX 4070 report
RTX 4070 Super8-bit20K37 tok/s0.5 sPDF: Lift on RTX 4070 Super report
RTX 50708-bit20K49 tok/s0.6 sPDF: Lift on RTX 5070 report
RTX 4060 Ti 16 GB8-bit144K21 tok/s0.8 sPDF: Lift on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit144K33 tok/s0.7 sPDF: Lift on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit144K29 tok/s0.8 sPDF: Lift on Arc A770 16 GB report
Radeon RX 9070 XT8-bit144K39 tok/s0.5 sPDF: Lift on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit144K49 tok/s0.4 sPDF: Lift on RTX 4070 Ti Super report
RTX 4080 Super8-bit144K53 tok/s0.3 sPDF: Lift on RTX 4080 Super report
RTX 5070 Ti8-bit144K64 tok/s0.4 sPDF: Lift on RTX 5070 Ti report
RTX 50808-bit144K68 tok/s0.3 sPDF: Lift on RTX 5080 report
Radeon RX 7900 XTX8-bit256K57 tok/s0.8 sPDF: Lift on Radeon RX 7900 XTX report
RTX 30908-bit256K66 tok/s0.5 sPDF: Lift on RTX 3090 report
RTX 40908-bit256K71 tok/s0.2 sPDF: Lift on RTX 4090 report
Radeon AI PRO R97008-bit256K39 tok/s0.5 sPDF: Lift on Radeon AI PRO R9700 report
RTX 50908-bit256K120 tok/s0.2 sPDF: Lift on RTX 5090 report
2× RTX 30908-bit256K66 tok/s0.3 sPDF: Lift on 2× RTX 3090 report
2× RTX 40908-bit256K71 tok/s0.1 sPDF: Lift on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit256K93 tok/s0.3 sPDF: Lift on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit256K68 tok/s0.2 sPDF: Lift on RTX 6000 Ada report
L40S8-bit256K62 tok/s0.3 sPDF: Lift on L40S report
2× RTX 50908-bit256K120 tok/s0.1 sPDF: Lift on 2× RTX 5090 report
4× RTX 30908-bit256K66 tok/s0.1 sPDF: Lift on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit256K120 tok/s0.1 sPDF: Lift on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB6-bit56K13 tok/s11 sPDF: Lift on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit144K10 tok/s11 sPDF: Lift on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit256K10 tok/s11 sPDF: Lift on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit144K23 tok/s5.4 sPDF: Lift on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit256K23 tok/s5.4 sPDF: Lift on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit256K23 tok/s5.4 sPDF: Lift on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit256K40 tok/s2.7 sPDF: Lift on M4 Max 128 GB report
M5 Max 128 GB8-bit256K44 tok/s0.8 sPDF: Lift on M5 Max 128 GB report
M3 Ultra 96 GB8-bit256K50 tok/s1.5 sPDF: Lift on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit256K50 tok/s1.5 sPDF: Lift on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit256K50 tok/s1.5 sPDF: Lift on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit256K69 tok/s0.5 sPDF: Lift on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit256K16 tok/s1.6 sPDF: Lift on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit256K16 tok/s1.6 sPDF: Lift on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit256K22 tok/s0.5 sPDF: Lift on RTX Spark report
DGX Spark8-bit256K20 tok/s0.5 sPDF: Lift on DGX Spark report
Dual DGX Spark8-bit256K20 tok/s0.3 sPDF: Lift on Dual DGX Spark report
DGX Station GB3008-bit256K355 tok/s0.1 sPDF: Lift on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit256K120 tok/s0.1 sPDF: Lift on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit256K120 tok/s0.1 sPDF: Lift on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit256K120 tok/s0.1 sPDF: Lift on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit256K135 tok/s0.3 sPDF: Lift on A100 80 GB SXM report
H100 SXM8-bit256K204 tok/s0.1 sPDF: Lift on H100 SXM report
4× H200 NVL8-bit256K270 tok/s0.1 sPDF: Lift on 4× H200 NVL report
8× H200 NVL8-bit256K270 tok/s0.1 sPDF: Lift 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 prithivMLmods/lift-GGUF
vLLM
vllm serve datalab-to/lift
MLX (Mac)
mlx_lm.chat --model mlx-community/lift-oQ4

Questions

How much VRAM does Lift need?

Lift needs 5.5 GB for its weights at 4-bit, 6.8 GB with an 8K context and 7.5 GB with 32K. At 8-bit, 10.9 GB.

What is the minimum VRAM for Lift?

About 5.7 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 Lift?

9.7B parameters: 18.0 GB of weights at 16-bit, 5.5 GB at 4-bit.

Can Lift run on an RTX 4090?

Yes. At 8-bit it holds 256K of context and writes about 71 tokens a second.

More models