Lift VRAM requirements
| 4-bit weights: 5.5 GB | With 8K context: 6.8 GB | With 32K context: 7.5 GB | 8-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
| Precision | Weights | With 8K | With 32K | With 128K |
|---|---|---|---|---|
| 16-bit | 18.0 GB | 19.3 GB | 20.1 GB | 23.1 GB |
| 8-bit | 9.6 GB | 10.9 GB | 11.6 GB | 14.6 GB |
| 6-bit | 7.4 GB | 8.7 GB | 9.5 GB | 12.5 GB |
| 5-bit | 6.4 GB | 7.7 GB | 8.5 GB | 11.5 GB |
| 4-bit | 5.5 GB | 6.8 GB | 7.5 GB | 10.5 GB |
| 3-bit | 4.4 GB | 5.7 GB | 6.5 GB | 9.5 GB |
| 2-bit | 3.4 GB | 4.7 GB | 5.5 GB | 8.5 GB |
| 1.6-bit | 1.9 GB | 3.2 GB | 4.0 GB | 7.0 GB |
One conversation, 16-bit context cache, runtime overhead included.
GPUs and Macs that run Lift
| Machine | Best fit | Context | Speed | First reply | Report |
|---|---|---|---|---|---|
| RTX 4060 | 4-bit | 32K | 35 tok/s | 1.1 s | PDF: Lift on RTX 4060 report |
| RTX 5060 | 4-bit | 32K | 56 tok/s | 0.9 s | PDF: Lift on RTX 5060 report |
| RTX 3070 | 4-bit | 32K | 56 tok/s | 0.8 s | PDF: Lift on RTX 3070 report |
| RTX 3080 10 GB | 6-bit | 32K | 70 tok/s | 0.6 s | PDF: Lift on RTX 3080 10 GB report |
| RTX 3060 12 GB | 8-bit | 20K | 27 tok/s | 1.3 s | PDF: Lift on RTX 3060 12 GB report |
| Arc B580 | 8-bit | 20K | 24 tok/s | 0.9 s | PDF: Lift on Arc B580 report |
| RTX 4070 | 8-bit | 20K | 37 tok/s | 0.6 s | PDF: Lift on RTX 4070 report |
| RTX 4070 Super | 8-bit | 20K | 37 tok/s | 0.5 s | PDF: Lift on RTX 4070 Super report |
| RTX 5070 | 8-bit | 20K | 49 tok/s | 0.6 s | PDF: Lift on RTX 5070 report |
| RTX 4060 Ti 16 GB | 8-bit | 144K | 21 tok/s | 0.8 s | PDF: Lift on RTX 4060 Ti 16 GB report |
| RTX 5060 Ti 16 GB | 8-bit | 144K | 33 tok/s | 0.7 s | PDF: Lift on RTX 5060 Ti 16 GB report |
| Arc A770 16 GB | 8-bit | 144K | 29 tok/s | 0.8 s | PDF: Lift on Arc A770 16 GB report |
| Radeon RX 9070 XT | 8-bit | 144K | 39 tok/s | 0.5 s | PDF: Lift on Radeon RX 9070 XT report |
| RTX 4070 Ti Super | 8-bit | 144K | 49 tok/s | 0.4 s | PDF: Lift on RTX 4070 Ti Super report |
| RTX 4080 Super | 8-bit | 144K | 53 tok/s | 0.3 s | PDF: Lift on RTX 4080 Super report |
| RTX 5070 Ti | 8-bit | 144K | 64 tok/s | 0.4 s | PDF: Lift on RTX 5070 Ti report |
| RTX 5080 | 8-bit | 144K | 68 tok/s | 0.3 s | PDF: Lift on RTX 5080 report |
| Radeon RX 7900 XTX | 8-bit | 256K | 57 tok/s | 0.8 s | PDF: Lift on Radeon RX 7900 XTX report |
| RTX 3090 | 8-bit | 256K | 66 tok/s | 0.5 s | PDF: Lift on RTX 3090 report |
| RTX 4090 | 8-bit | 256K | 71 tok/s | 0.2 s | PDF: Lift on RTX 4090 report |
| Radeon AI PRO R9700 | 8-bit | 256K | 39 tok/s | 0.5 s | PDF: Lift on Radeon AI PRO R9700 report |
| RTX 5090 | 8-bit | 256K | 120 tok/s | 0.2 s | PDF: Lift on RTX 5090 report |
| 2× RTX 3090 | 8-bit | 256K | 66 tok/s | 0.3 s | PDF: Lift on 2× RTX 3090 report |
| 2× RTX 4090 | 8-bit | 256K | 71 tok/s | 0.1 s | PDF: Lift on 2× RTX 4090 report |
| RTX PRO 5000 Blackwell 48 GB | 8-bit | 256K | 93 tok/s | 0.3 s | PDF: Lift on RTX PRO 5000 Blackwell 48 GB report |
| RTX 6000 Ada | 8-bit | 256K | 68 tok/s | 0.2 s | PDF: Lift on RTX 6000 Ada report |
| L40S | 8-bit | 256K | 62 tok/s | 0.3 s | PDF: Lift on L40S report |
| 2× RTX 5090 | 8-bit | 256K | 120 tok/s | 0.1 s | PDF: Lift on 2× RTX 5090 report |
| 4× RTX 3090 | 8-bit | 256K | 66 tok/s | 0.1 s | PDF: Lift on 4× RTX 3090 report |
| RTX PRO 6000 Blackwell | 8-bit | 256K | 120 tok/s | 0.1 s | PDF: Lift on RTX PRO 6000 Blackwell report |
| Mac mini M4 16 GB | 6-bit | 56K | 13 tok/s | 11 s | PDF: Lift on Mac mini M4 16 GB report |
| Mac mini M4 24 GB | 8-bit | 144K | 10 tok/s | 11 s | PDF: Lift on Mac mini M4 24 GB report |
| Mac mini M4 32 GB | 8-bit | 256K | 10 tok/s | 11 s | PDF: Lift on Mac mini M4 32 GB report |
| Mac mini M4 Pro 24 GB | 8-bit | 144K | 23 tok/s | 5.4 s | PDF: Lift on Mac mini M4 Pro 24 GB report |
| Mac mini M4 Pro 48 GB | 8-bit | 256K | 23 tok/s | 5.4 s | PDF: Lift on Mac mini M4 Pro 48 GB report |
| Mac mini M4 Pro 64 GB | 8-bit | 256K | 23 tok/s | 5.4 s | PDF: Lift on Mac mini M4 Pro 64 GB report |
| M4 Max 128 GB | 8-bit | 256K | 40 tok/s | 2.7 s | PDF: Lift on M4 Max 128 GB report |
| M5 Max 128 GB | 8-bit | 256K | 44 tok/s | 0.8 s | PDF: Lift on M5 Max 128 GB report |
| M3 Ultra 96 GB | 8-bit | 256K | 50 tok/s | 1.5 s | PDF: Lift on M3 Ultra 96 GB report |
| M3 Ultra 256 GB | 8-bit | 256K | 50 tok/s | 1.5 s | PDF: Lift on M3 Ultra 256 GB report |
| M3 Ultra 512 GB | 8-bit | 256K | 50 tok/s | 1.5 s | PDF: Lift on M3 Ultra 512 GB report |
| M5 Ultra 512 GB | 8-bit | 256K | 69 tok/s | 0.5 s | PDF: Lift on M5 Ultra 512 GB report |
| Ryzen AI Max+ 395 64 GB | 8-bit | 256K | 16 tok/s | 1.6 s | PDF: Lift on Ryzen AI Max+ 395 64 GB report |
| Ryzen AI Max+ 395 128 GB | 8-bit | 256K | 16 tok/s | 1.6 s | PDF: Lift on Ryzen AI Max+ 395 128 GB report |
| RTX Spark | 8-bit | 256K | 22 tok/s | 0.5 s | PDF: Lift on RTX Spark report |
| DGX Spark | 8-bit | 256K | 20 tok/s | 0.5 s | PDF: Lift on DGX Spark report |
| Dual DGX Spark | 8-bit | 256K | 20 tok/s | 0.3 s | PDF: Lift on Dual DGX Spark report |
| DGX Station GB300 | 8-bit | 256K | 355 tok/s | 0.1 s | PDF: Lift on DGX Station GB300 report |
| 2× RTX PRO 6000 Blackwell | 8-bit | 256K | 120 tok/s | 0.1 s | PDF: Lift on 2× RTX PRO 6000 Blackwell report |
| 4× RTX PRO 6000 Blackwell | 8-bit | 256K | 120 tok/s | 0.1 s | PDF: Lift on 4× RTX PRO 6000 Blackwell report |
| 8× RTX PRO 6000 Blackwell | 8-bit | 256K | 120 tok/s | 0.1 s | PDF: Lift on 8× RTX PRO 6000 Blackwell report |
| A100 80 GB SXM | 8-bit | 256K | 135 tok/s | 0.3 s | PDF: Lift on A100 80 GB SXM report |
| H100 SXM | 8-bit | 256K | 204 tok/s | 0.1 s | PDF: Lift on H100 SXM report |
| 4× H200 NVL | 8-bit | 256K | 270 tok/s | 0.1 s | PDF: Lift on 4× H200 NVL report |
| 8× H200 NVL | 8-bit | 256K | 270 tok/s | 0.1 s | PDF: 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.