Skip to content

Sarashina2.2 3B Instruct V0.1 VRAM requirements

How much VRAM Sarashina2.2 3B Instruct V0.1 needs
4-bit weights: 1.9 GBWith 8K context: 3.9 GB8-bit, 8K context: 5.3 GB
Parameters
3.4B
Longest context
8K
Released
Feb 2025
License
MIT
Strongest at
Not rated

How much VRAM Sarashina2.2 3B Instruct V0.1 needs

Memory Sarashina2.2 3B Instruct V0.1 needs at each precision and context length
PrecisionWeightsWith 8K
16-bit6.3 GB8.2 GB
8-bit3.3 GB5.3 GB
6-bit2.6 GB4.5 GB
5-bit2.2 GB4.2 GB
4-bit1.9 GB3.9 GB
3-bit1.5 GB3.5 GB
2-bit1.2 GB3.1 GB
1.6-bit660 MB2.6 GB

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

GPUs and Macs that run Sarashina2.2 3B Instruct V0.1

The best way Sarashina2.2 3B Instruct V0.1 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit8K55 tok/s0.4 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 4060 report
RTX 50608-bit8K87 tok/s0.3 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 5060 report
RTX 30708-bit8K87 tok/s0.3 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 3070 report
RTX 3080 10 GB8-bit8K140 tok/s0.2 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 3080 10 GB report
RTX 3060 12 GB8-bit8K71 tok/s0.5 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 3060 12 GB report
Arc B5808-bit8K61 tok/s0.4 sPDF: Sarashina2.2 3B Instruct V0.1 on Arc B580 report
RTX 40708-bit8K97 tok/s0.2 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 4070 report
RTX 4070 Super8-bit8K97 tok/s0.2 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 4070 Super report
RTX 50708-bit8K126 tok/s0.2 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 5070 report
RTX 4060 Ti 16 GB8-bit8K58 tok/s0.3 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit8K87 tok/s0.3 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit8K73 tok/s0.3 sPDF: Sarashina2.2 3B Instruct V0.1 on Arc A770 16 GB report
Radeon RX 9070 XT8-bit8K98 tok/s0.2 sPDF: Sarashina2.2 3B Instruct V0.1 on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit8K126 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 4070 Ti Super report
RTX 4080 Super8-bit8K136 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 4080 Super report
RTX 5070 Ti8-bit8K161 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 5070 Ti report
RTX 50808-bit8K171 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 5080 report
Radeon RX 7900 XTX8-bit8K137 tok/s0.3 sPDF: Sarashina2.2 3B Instruct V0.1 on Radeon RX 7900 XTX report
RTX 30908-bit8K167 tok/s0.2 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 3090 report
RTX 40908-bit8K178 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 4090 report
Radeon AI PRO R97008-bit8K98 tok/s0.2 sPDF: Sarashina2.2 3B Instruct V0.1 on Radeon AI PRO R9700 report
RTX 50908-bit8K279 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 5090 report
2× RTX 30908-bit8K167 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on 2× RTX 3090 report
2× RTX 40908-bit8K178 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit8K225 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit8K171 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX 6000 Ada report
L40S8-bit8K156 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on L40S report
2× RTX 50908-bit8K279 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on 2× RTX 5090 report
4× RTX 30908-bit8K167 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit8K279 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit8K28 tok/s4.1 sPDF: Sarashina2.2 3B Instruct V0.1 on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit8K28 tok/s4.1 sPDF: Sarashina2.2 3B Instruct V0.1 on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit8K28 tok/s4.1 sPDF: Sarashina2.2 3B Instruct V0.1 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit8K58 tok/s2.1 sPDF: Sarashina2.2 3B Instruct V0.1 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit8K58 tok/s2.1 sPDF: Sarashina2.2 3B Instruct V0.1 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit8K58 tok/s2.1 sPDF: Sarashina2.2 3B Instruct V0.1 on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit8K94 tok/s1.0 sPDF: Sarashina2.2 3B Instruct V0.1 on M4 Max 128 GB report
M5 Max 128 GB8-bit8K103 tok/s0.3 sPDF: Sarashina2.2 3B Instruct V0.1 on M5 Max 128 GB report
M3 Ultra 96 GB8-bit8K114 tok/s0.6 sPDF: Sarashina2.2 3B Instruct V0.1 on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit8K114 tok/s0.6 sPDF: Sarashina2.2 3B Instruct V0.1 on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit8K114 tok/s0.6 sPDF: Sarashina2.2 3B Instruct V0.1 on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit8K148 tok/s0.2 sPDF: Sarashina2.2 3B Instruct V0.1 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit8K43 tok/s0.6 sPDF: Sarashina2.2 3B Instruct V0.1 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit8K43 tok/s0.6 sPDF: Sarashina2.2 3B Instruct V0.1 on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit8K60 tok/s0.2 sPDF: Sarashina2.2 3B Instruct V0.1 on RTX Spark report
DGX Spark8-bit8K55 tok/s0.2 sPDF: Sarashina2.2 3B Instruct V0.1 on DGX Spark report
Dual DGX Spark8-bit8K55 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on Dual DGX Spark report
DGX Station GB3008-bit8K617 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit8K279 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit8K279 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit8K279 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit8K306 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on A100 80 GB SXM report
H100 SXM8-bit8K423 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on H100 SXM report
4× H200 NVL8-bit8K516 tok/s0.1 sPDF: Sarashina2.2 3B Instruct V0.1 on 4× H200 NVL report
8× H200 NVL8-bit8K516 tok/s0.1 sPDF: Sarashina2.2 3B Instruct 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 mmnga/sarashina2.2-3b-instruct-v0.1-gguf
vLLM
vllm serve sbintuitions/sarashina2.2-3b-instruct-v0.1
MLX (Mac)
mlx_lm.chat --model mlx-community/sarashina2.2-3b-instruct-v0.1-4bit

Questions

How much VRAM does Sarashina2.2 3B Instruct V0.1 need?

Sarashina2.2 3B Instruct V0.1 needs 1.9 GB for its weights at 4-bit, 3.9 GB with an 8K context and 3.9 GB with 8K. At 8-bit, 5.3 GB.

What is the minimum VRAM for Sarashina2.2 3B Instruct V0.1?

About 3.5 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 Sarashina2.2 3B Instruct V0.1?

3.4B parameters: 6.3 GB of weights at 16-bit, 1.9 GB at 4-bit.

Can Sarashina2.2 3B Instruct V0.1 run on an RTX 4090?

Yes. At 8-bit it holds 8K of context and writes about 178 tokens a second.

More models