Skip to content

TinyLlama 1.1B Step 50K 105B VRAM requirements

How much VRAM TinyLlama 1.1B Step 50K 105B needs
4-bit weights: 636 MBWith 2K context: 1.2 GB8-bit, 2K context: 1.7 GB
Parameters
1.1B
Longest context
2K
Released
Sep 2023
License
Apache 2.0
Strongest at
Not rated

How much VRAM TinyLlama 1.1B Step 50K 105B needs

Memory TinyLlama 1.1B Step 50K 105B needs at each precision and context length
PrecisionWeightsWith 2K
16-bit2.0 GB2.7 GB
8-bit1.1 GB1.7 GB
6-bit860 MB1.5 GB
5-bit746 MB1.3 GB
4-bit636 MB1.2 GB
3-bit513 MB1.1 GB
2-bit393 MB1.0 GB
1.6-bit216 MB848 MB

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

GPUs and Macs that run TinyLlama 1.1B Step 50K 105B

The best way TinyLlama 1.1B Step 50K 105B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit2K155 tok/s0.2 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 4060 report
RTX 50608-bit2K240 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 5060 report
RTX 30708-bit2K240 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 3070 report
RTX 3080 10 GB8-bit2K367 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 3080 10 GB report
RTX 3060 12 GB8-bit2K199 tok/s0.2 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 3060 12 GB report
Arc B5808-bit2K163 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on Arc B580 report
RTX 40708-bit2K265 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 4070 report
RTX 4070 Super8-bit2K265 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 4070 Super report
RTX 50708-bit2K334 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 5070 report
RTX 4060 Ti 16 GB8-bit2K164 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit2K240 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit2K191 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on Arc A770 16 GB report
Radeon RX 9070 XT8-bit2K252 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit2K334 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 4070 Ti Super report
RTX 4080 Super8-bit2K358 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 4080 Super report
RTX 5070 Ti8-bit2K415 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 5070 Ti report
RTX 50808-bit2K436 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 5080 report
Radeon RX 7900 XTX8-bit2K332 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on Radeon RX 7900 XTX report
RTX 30908-bit2K428 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 3090 report
RTX 40908-bit2K451 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 4090 report
Radeon AI PRO R97008-bit2K252 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on Radeon AI PRO R9700 report
RTX 50908-bit2K651 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 5090 report
2× RTX 30908-bit2K428 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on 2× RTX 3090 report
2× RTX 40908-bit2K451 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit2K547 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit2K436 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX 6000 Ada report
L40S8-bit2K404 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on L40S report
2× RTX 50908-bit2K651 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on 2× RTX 5090 report
4× RTX 30908-bit2K428 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit2K651 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit2K77 tok/s1.5 sPDF: TinyLlama 1.1B Step 50K 105B on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit2K77 tok/s1.5 sPDF: TinyLlama 1.1B Step 50K 105B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit2K77 tok/s1.5 sPDF: TinyLlama 1.1B Step 50K 105B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit2K150 tok/s0.7 sPDF: TinyLlama 1.1B Step 50K 105B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit2K150 tok/s0.7 sPDF: TinyLlama 1.1B Step 50K 105B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit2K150 tok/s0.7 sPDF: TinyLlama 1.1B Step 50K 105B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit2K225 tok/s0.4 sPDF: TinyLlama 1.1B Step 50K 105B on M4 Max 128 GB report
M5 Max 128 GB8-bit2K241 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit2K261 tok/s0.2 sPDF: TinyLlama 1.1B Step 50K 105B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit2K261 tok/s0.2 sPDF: TinyLlama 1.1B Step 50K 105B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit2K261 tok/s0.2 sPDF: TinyLlama 1.1B Step 50K 105B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit2K316 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit2K121 tok/s0.2 sPDF: TinyLlama 1.1B Step 50K 105B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit2K121 tok/s0.2 sPDF: TinyLlama 1.1B Step 50K 105B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit2K170 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on RTX Spark report
DGX Spark8-bit2K156 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on DGX Spark report
Dual DGX Spark8-bit2K156 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on Dual DGX Spark report
DGX Station GB3008-bit2K1,135 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit2K651 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit2K651 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit2K651 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit2K699 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on A100 80 GB SXM report
H100 SXM8-bit2K886 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on H100 SXM report
4× H200 NVL8-bit2K1,013 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B on 4× H200 NVL report
8× H200 NVL8-bit2K1,013 tok/s0.1 sPDF: TinyLlama 1.1B Step 50K 105B 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 tensorblock/TinyLlama-1.1B-step-50K-105b-GGUF
vLLM
vllm serve TinyLlama/TinyLlama-1.1B-step-50K-105b

Questions

How much VRAM does TinyLlama 1.1B Step 50K 105B need?

TinyLlama 1.1B Step 50K 105B needs 636 MB for its weights at 4-bit, 1.2 GB with an 2K context and 1.2 GB with 2K. At 8-bit, 1.7 GB.

What is the minimum VRAM for TinyLlama 1.1B Step 50K 105B?

About 1.1 GB at 3-bit with an 2K 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 TinyLlama 1.1B Step 50K 105B?

1.1B parameters: 2.0 GB of weights at 16-bit, 636 MB at 4-bit.

Can TinyLlama 1.1B Step 50K 105B run on an RTX 4090?

Yes. At 8-bit it holds 2K of context and writes about 451 tokens a second.

More models