Skip to content

H2o Danube3 500M Chat VRAM requirements

How much VRAM H2o Danube3 500M Chat needs
4-bit weights: 297 MBWith 8K context: 1.2 GB8-bit, 8K context: 1.4 GB
Parameters
514M
Longest context
8K
Released
Jul 2024
License
Apache 2.0
Strongest at
Not rated

How much VRAM H2o Danube3 500M Chat needs

Memory H2o Danube3 500M Chat needs at each precision and context length
PrecisionWeightsWith 8K
16-bit1.0 GB1.9 GB
8-bit520 MB1.4 GB
6-bit402 MB1.3 GB
5-bit348 MB1.3 GB
4-bit297 MB1.2 GB
3-bit239 MB1.2 GB
2-bit184 MB1.1 GB
1.6-bit101 MB1.0 GB

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

GPUs and Macs that run H2o Danube3 500M Chat

The best way H2o Danube3 500M Chat runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit8K326 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 4060 report
RTX 50608-bit8K487 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 5060 report
RTX 30708-bit8K487 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 3070 report
RTX 3080 10 GB8-bit8K710 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 3080 10 GB report
RTX 3060 12 GB8-bit8K410 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 3060 12 GB report
Arc B5808-bit8K320 tok/s0.1 sPDF: H2o Danube3 500M Chat on Arc B580 report
RTX 40708-bit8K532 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 4070 report
RTX 4070 Super8-bit8K532 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 4070 Super report
RTX 50708-bit8K654 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 5070 report
RTX 4060 Ti 16 GB8-bit8K342 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit8K487 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit8K367 tok/s0.1 sPDF: H2o Danube3 500M Chat on Arc A770 16 GB report
Radeon RX 9070 XT8-bit8K480 tok/s0.1 sPDF: H2o Danube3 500M Chat on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit8K654 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 4070 Ti Super report
RTX 4080 Super8-bit8K695 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 4080 Super report
RTX 5070 Ti8-bit8K789 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 5070 Ti report
RTX 50808-bit8K824 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 5080 report
Radeon RX 7900 XTX8-bit8K604 tok/s0.1 sPDF: H2o Danube3 500M Chat on Radeon RX 7900 XTX report
RTX 30908-bit8K811 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 3090 report
RTX 40908-bit8K848 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 4090 report
Radeon AI PRO R97008-bit8K480 tok/s0.1 sPDF: H2o Danube3 500M Chat on Radeon AI PRO R9700 report
RTX 50908-bit8K1,145 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 5090 report
2× RTX 30908-bit8K811 tok/s0.1 sPDF: H2o Danube3 500M Chat on 2× RTX 3090 report
2× RTX 40908-bit8K848 tok/s0.1 sPDF: H2o Danube3 500M Chat on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit8K996 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit8K824 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX 6000 Ada report
L40S8-bit8K772 tok/s0.1 sPDF: H2o Danube3 500M Chat on L40S report
2× RTX 50908-bit8K1,145 tok/s0.1 sPDF: H2o Danube3 500M Chat on 2× RTX 5090 report
4× RTX 30908-bit8K811 tok/s0.1 sPDF: H2o Danube3 500M Chat on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit8K1,145 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit8K159 tok/s0.6 sPDF: H2o Danube3 500M Chat on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit8K159 tok/s0.6 sPDF: H2o Danube3 500M Chat on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit8K159 tok/s0.6 sPDF: H2o Danube3 500M Chat on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit8K287 tok/s0.3 sPDF: H2o Danube3 500M Chat on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit8K287 tok/s0.3 sPDF: H2o Danube3 500M Chat on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit8K287 tok/s0.3 sPDF: H2o Danube3 500M Chat on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit8K402 tok/s0.2 sPDF: H2o Danube3 500M Chat on M4 Max 128 GB report
M5 Max 128 GB8-bit8K425 tok/s0.1 sPDF: H2o Danube3 500M Chat on M5 Max 128 GB report
M3 Ultra 96 GB8-bit8K453 tok/s0.1 sPDF: H2o Danube3 500M Chat on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit8K453 tok/s0.1 sPDF: H2o Danube3 500M Chat on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit8K453 tok/s0.1 sPDF: H2o Danube3 500M Chat on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit8K523 tok/s0.1 sPDF: H2o Danube3 500M Chat on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit8K249 tok/s0.1 sPDF: H2o Danube3 500M Chat on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit8K249 tok/s0.1 sPDF: H2o Danube3 500M Chat on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit8K353 tok/s0.1 sPDF: H2o Danube3 500M Chat on RTX Spark report
DGX Spark8-bit8K327 tok/s0.1 sPDF: H2o Danube3 500M Chat on DGX Spark report
Dual DGX Spark8-bit8K327 tok/s0.1 sPDF: H2o Danube3 500M Chat on Dual DGX Spark report
DGX Station GB3008-bit8K1,726 tok/s0.1 sPDF: H2o Danube3 500M Chat on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit8K1,145 tok/s0.1 sPDF: H2o Danube3 500M Chat on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit8K1,145 tok/s0.1 sPDF: H2o Danube3 500M Chat on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit8K1,145 tok/s0.1 sPDF: H2o Danube3 500M Chat on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit8K1,211 tok/s0.1 sPDF: H2o Danube3 500M Chat on A100 80 GB SXM report
H100 SXM8-bit8K1,448 tok/s0.1 sPDF: H2o Danube3 500M Chat on H100 SXM report
4× H200 NVL8-bit8K1,595 tok/s0.1 sPDF: H2o Danube3 500M Chat on 4× H200 NVL report
8× H200 NVL8-bit8K1,595 tok/s0.1 sPDF: H2o Danube3 500M Chat 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 bartowski/h2o-danube3-500m-chat-GGUF
vLLM
vllm serve h2oai/h2o-danube3-500m-chat
MLX (Mac)
mlx_lm.chat --model mlx-community/h2o-danube3-500m-chat

Questions

How much VRAM does H2o Danube3 500M Chat need?

H2o Danube3 500M Chat needs 297 MB for its weights at 4-bit, 1.2 GB with an 8K context and 1.2 GB with 8K. At 8-bit, 1.4 GB.

What is the minimum VRAM for H2o Danube3 500M Chat?

About 1.2 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 H2o Danube3 500M Chat?

514M parameters: 1.0 GB of weights at 16-bit, 297 MB at 4-bit.

Can H2o Danube3 500M Chat run on an RTX 4090?

Yes. At 8-bit it holds 8K of context and writes about 848 tokens a second.

More models