Skip to content

Mistral 7B V0.1 VRAM requirements

How much VRAM Mistral 7B V0.1 needs
4-bit weights: 4.1 GBWith 8K context: 5.2 GBWith 32K context: 5.2 GB8-bit, 8K context: 8.3 GB
Parameters
7.2B
Longest context
32K
Released
Sep 2023
License
Apache 2.0
Strongest at
Not rated

How much VRAM Mistral 7B V0.1 needs

Memory Mistral 7B V0.1 needs at each precision and context length
PrecisionWeightsWith 8KWith 32K
16-bit13.5 GB14.6 GB14.6 GB
8-bit7.2 GB8.3 GB8.3 GB
6-bit5.5 GB6.7 GB6.7 GB
5-bit4.8 GB5.9 GB5.9 GB
4-bit4.1 GB5.2 GB5.2 GB
3-bit3.3 GB4.4 GB4.4 GB
2-bit2.5 GB3.7 GB3.7 GB
1.6-bit1.4 GB2.5 GB2.5 GB

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

GPUs and Macs that run Mistral 7B V0.1

The best way Mistral 7B V0.1 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40606-bit32K32 tok/s1.1 sPDF: Mistral 7B V0.1 on RTX 4060 report
RTX 50606-bit32K51 tok/s0.8 sPDF: Mistral 7B V0.1 on RTX 5060 report
RTX 30706-bit32K51 tok/s0.8 sPDF: Mistral 7B V0.1 on RTX 3070 report
RTX 3080 10 GB8-bit32K66 tok/s0.5 sPDF: Mistral 7B V0.1 on RTX 3080 10 GB report
RTX 3060 12 GB8-bit32K32 tok/s1.3 sPDF: Mistral 7B V0.1 on RTX 3060 12 GB report
Arc B5808-bit32K29 tok/s0.9 sPDF: Mistral 7B V0.1 on Arc B580 report
RTX 40708-bit32K45 tok/s0.6 sPDF: Mistral 7B V0.1 on RTX 4070 report
RTX 4070 Super8-bit32K45 tok/s0.5 sPDF: Mistral 7B V0.1 on RTX 4070 Super report
RTX 50708-bit32K59 tok/s0.5 sPDF: Mistral 7B V0.1 on RTX 5070 report
RTX 4060 Ti 16 GB8-bit32K26 tok/s0.7 sPDF: Mistral 7B V0.1 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit32K40 tok/s0.7 sPDF: Mistral 7B V0.1 on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit32K35 tok/s0.8 sPDF: Mistral 7B V0.1 on Arc A770 16 GB report
Radeon RX 9070 XT8-bit32K47 tok/s0.5 sPDF: Mistral 7B V0.1 on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit32K59 tok/s0.4 sPDF: Mistral 7B V0.1 on RTX 4070 Ti Super report
RTX 4080 Super8-bit32K64 tok/s0.3 sPDF: Mistral 7B V0.1 on RTX 4080 Super report
RTX 5070 Ti8-bit32K77 tok/s0.4 sPDF: Mistral 7B V0.1 on RTX 5070 Ti report
RTX 50808-bit32K82 tok/s0.3 sPDF: Mistral 7B V0.1 on RTX 5080 report
Radeon RX 7900 XTX8-bit32K68 tok/s0.7 sPDF: Mistral 7B V0.1 on Radeon RX 7900 XTX report
RTX 30908-bit32K80 tok/s0.5 sPDF: Mistral 7B V0.1 on RTX 3090 report
RTX 40908-bit32K86 tok/s0.2 sPDF: Mistral 7B V0.1 on RTX 4090 report
Radeon AI PRO R97008-bit32K47 tok/s0.5 sPDF: Mistral 7B V0.1 on Radeon AI PRO R9700 report
RTX 50908-bit32K143 tok/s0.2 sPDF: Mistral 7B V0.1 on RTX 5090 report
2× RTX 30908-bit32K80 tok/s0.3 sPDF: Mistral 7B V0.1 on 2× RTX 3090 report
2× RTX 40908-bit32K86 tok/s0.1 sPDF: Mistral 7B V0.1 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit32K111 tok/s0.2 sPDF: Mistral 7B V0.1 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit32K82 tok/s0.2 sPDF: Mistral 7B V0.1 on RTX 6000 Ada report
L40S8-bit32K74 tok/s0.3 sPDF: Mistral 7B V0.1 on L40S report
2× RTX 50908-bit32K143 tok/s0.1 sPDF: Mistral 7B V0.1 on 2× RTX 5090 report
4× RTX 30908-bit32K80 tok/s0.1 sPDF: Mistral 7B V0.1 on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit32K143 tok/s0.1 sPDF: Mistral 7B V0.1 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit32K13 tok/s10 sPDF: Mistral 7B V0.1 on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit32K13 tok/s10 sPDF: Mistral 7B V0.1 on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit32K13 tok/s10 sPDF: Mistral 7B V0.1 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit32K28 tok/s5.1 sPDF: Mistral 7B V0.1 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit32K28 tok/s5.1 sPDF: Mistral 7B V0.1 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit32K28 tok/s5.1 sPDF: Mistral 7B V0.1 on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit32K48 tok/s2.6 sPDF: Mistral 7B V0.1 on M4 Max 128 GB report
M5 Max 128 GB8-bit32K53 tok/s0.7 sPDF: Mistral 7B V0.1 on M5 Max 128 GB report
M3 Ultra 96 GB8-bit32K60 tok/s1.4 sPDF: Mistral 7B V0.1 on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit32K60 tok/s1.4 sPDF: Mistral 7B V0.1 on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit32K60 tok/s1.4 sPDF: Mistral 7B V0.1 on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit32K82 tok/s0.5 sPDF: Mistral 7B V0.1 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit32K20 tok/s1.5 sPDF: Mistral 7B V0.1 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit32K20 tok/s1.5 sPDF: Mistral 7B V0.1 on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit32K27 tok/s0.5 sPDF: Mistral 7B V0.1 on RTX Spark report
DGX Spark8-bit32K25 tok/s0.5 sPDF: Mistral 7B V0.1 on DGX Spark report
Dual DGX Spark8-bit32K25 tok/s0.3 sPDF: Mistral 7B V0.1 on Dual DGX Spark report
DGX Station GB3008-bit32K403 tok/s0.1 sPDF: Mistral 7B V0.1 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit32K143 tok/s0.1 sPDF: Mistral 7B V0.1 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit32K143 tok/s0.1 sPDF: Mistral 7B V0.1 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit32K143 tok/s0.1 sPDF: Mistral 7B V0.1 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit32K160 tok/s0.3 sPDF: Mistral 7B V0.1 on A100 80 GB SXM report
H100 SXM8-bit32K239 tok/s0.1 sPDF: Mistral 7B V0.1 on H100 SXM report
4× H200 NVL8-bit32K312 tok/s0.1 sPDF: Mistral 7B V0.1 on 4× H200 NVL report
8× H200 NVL8-bit32K312 tok/s0.1 sPDF: Mistral 7B V0.1 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 NousResearch/Hermes-2-Pro-Mistral-7B-GGUF
vLLM
vllm serve mistralai/Mistral-7B-v0.1

Questions

How much VRAM does Mistral 7B V0.1 need?

Mistral 7B V0.1 needs 4.1 GB for its weights at 4-bit, 5.2 GB with an 8K context and 5.2 GB with 32K. At 8-bit, 8.3 GB.

What is the minimum VRAM for Mistral 7B V0.1?

About 4.4 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 Mistral 7B V0.1?

7.2B parameters: 13.5 GB of weights at 16-bit, 4.1 GB at 4-bit.

Can Mistral 7B V0.1 run on an RTX 4090?

Yes. At 8-bit it holds 32K of context and writes about 86 tokens a second.

More models