Skip to content

Spark X2.5 4B VRAM requirements

How much VRAM Spark X2.5 4B needs
4-bit weights: 2.3 GBWith 8K context: 3.4 GBWith 32K context: 4.3 GB8-bit, 8K context: 5.2 GB
Parameters
4.1B
Longest context
1M
Released
Aug 2026
License
Apache 2.0
Strongest at
Not rated

How much VRAM Spark X2.5 4B needs

Memory Spark X2.5 4B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit7.7 GB8.8 GB9.6 GB13.0 GB
8-bit4.1 GB5.2 GB6.0 GB9.4 GB
6-bit3.1 GB4.3 GB5.1 GB8.5 GB
5-bit2.7 GB3.8 GB4.7 GB8.1 GB
4-bit2.3 GB3.4 GB4.3 GB7.7 GB
3-bit1.9 GB3.0 GB3.8 GB7.2 GB
2-bit1.4 GB2.5 GB3.4 GB6.8 GB
1.6-bit809 MB1.9 GB2.7 GB6.1 GB

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

GPUs and Macs that run Spark X2.5 4B

The best way Spark X2.5 4B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit72K42 tok/s0.6 sPDF: Spark X2.5 4B on RTX 4060 report
RTX 50608-bit72K67 tok/s0.4 sPDF: Spark X2.5 4B on RTX 5060 report
RTX 30708-bit72K67 tok/s0.4 sPDF: Spark X2.5 4B on RTX 3070 report
RTX 3080 10 GB8-bit128K108 tok/s0.3 sPDF: Spark X2.5 4B on RTX 3080 10 GB report
RTX 3060 12 GB8-bit184K54 tok/s0.7 sPDF: Spark X2.5 4B on RTX 3060 12 GB report
Arc B5808-bit184K47 tok/s0.5 sPDF: Spark X2.5 4B on Arc B580 report
RTX 40708-bit184K74 tok/s0.3 sPDF: Spark X2.5 4B on RTX 4070 report
RTX 4070 Super8-bit184K74 tok/s0.2 sPDF: Spark X2.5 4B on RTX 4070 Super report
RTX 50708-bit184K96 tok/s0.3 sPDF: Spark X2.5 4B on RTX 5070 report
RTX 4060 Ti 16 GB8-bit288K44 tok/s0.4 sPDF: Spark X2.5 4B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit288K67 tok/s0.4 sPDF: Spark X2.5 4B on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit288K56 tok/s0.4 sPDF: Spark X2.5 4B on Arc A770 16 GB report
Radeon RX 9070 XT8-bit288K76 tok/s0.2 sPDF: Spark X2.5 4B on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit288K96 tok/s0.2 sPDF: Spark X2.5 4B on RTX 4070 Ti Super report
RTX 4080 Super8-bit288K105 tok/s0.2 sPDF: Spark X2.5 4B on RTX 4080 Super report
RTX 5070 Ti8-bit288K124 tok/s0.2 sPDF: Spark X2.5 4B on RTX 5070 Ti report
RTX 50808-bit288K132 tok/s0.1 sPDF: Spark X2.5 4B on RTX 5080 report
Radeon RX 7900 XTX8-bit480K107 tok/s0.4 sPDF: Spark X2.5 4B on Radeon RX 7900 XTX report
RTX 30908-bit480K129 tok/s0.2 sPDF: Spark X2.5 4B on RTX 3090 report
RTX 40908-bit480K138 tok/s0.1 sPDF: Spark X2.5 4B on RTX 4090 report
Radeon AI PRO R97008-bit704K76 tok/s0.2 sPDF: Spark X2.5 4B on Radeon AI PRO R9700 report
RTX 50908-bit704K219 tok/s0.1 sPDF: Spark X2.5 4B on RTX 5090 report
2× RTX 30908-bit1M129 tok/s0.1 sPDF: Spark X2.5 4B on 2× RTX 3090 report
2× RTX 40908-bit1M138 tok/s0.1 sPDF: Spark X2.5 4B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit1M175 tok/s0.1 sPDF: Spark X2.5 4B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit1M132 tok/s0.1 sPDF: Spark X2.5 4B on RTX 6000 Ada report
L40S8-bit1M120 tok/s0.1 sPDF: Spark X2.5 4B on L40S report
2× RTX 50908-bit1M219 tok/s0.1 sPDF: Spark X2.5 4B on 2× RTX 5090 report
4× RTX 30908-bit1M129 tok/s0.1 sPDF: Spark X2.5 4B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit1M219 tok/s0.1 sPDF: Spark X2.5 4B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit152K21 tok/s5.3 sPDF: Spark X2.5 4B on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit288K21 tok/s5.3 sPDF: Spark X2.5 4B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit416K21 tok/s5.3 sPDF: Spark X2.5 4B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit288K45 tok/s2.7 sPDF: Spark X2.5 4B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit832K45 tok/s2.7 sPDF: Spark X2.5 4B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit1M45 tok/s2.7 sPDF: Spark X2.5 4B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit1M74 tok/s1.3 sPDF: Spark X2.5 4B on M4 Max 128 GB report
M5 Max 128 GB8-bit1M81 tok/s0.4 sPDF: Spark X2.5 4B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit1M90 tok/s0.7 sPDF: Spark X2.5 4B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit1M90 tok/s0.7 sPDF: Spark X2.5 4B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit1M90 tok/s0.7 sPDF: Spark X2.5 4B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit1M118 tok/s0.2 sPDF: Spark X2.5 4B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit1M33 tok/s0.8 sPDF: Spark X2.5 4B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit1M33 tok/s0.8 sPDF: Spark X2.5 4B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit1M46 tok/s0.3 sPDF: Spark X2.5 4B on RTX Spark report
DGX Spark8-bit1M42 tok/s0.3 sPDF: Spark X2.5 4B on DGX Spark report
Dual DGX Spark8-bit1M42 tok/s0.2 sPDF: Spark X2.5 4B on Dual DGX Spark report
DGX Station GB3008-bit1M510 tok/s0.1 sPDF: Spark X2.5 4B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit1M219 tok/s0.1 sPDF: Spark X2.5 4B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit1M219 tok/s0.1 sPDF: Spark X2.5 4B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit1M219 tok/s0.1 sPDF: Spark X2.5 4B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit1M241 tok/s0.2 sPDF: Spark X2.5 4B on A100 80 GB SXM report
H100 SXM8-bit1M340 tok/s0.1 sPDF: Spark X2.5 4B on H100 SXM report
4× H200 NVL8-bit1M420 tok/s0.1 sPDF: Spark X2.5 4B on 4× H200 NVL report
8× H200 NVL8-bit1M420 tok/s0.1 sPDF: Spark X2.5 4B 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 peculiar-ragdoll/Sharp-Spark-X2.5-4B-GGUF
vLLM
vllm serve XHToken/Spark-X2.5-4B
MLX (Mac)
mlx_lm.chat --model mlx-community/Spark-X2.5-4B-OptiQ-4bit

Questions

How much VRAM does Spark X2.5 4B need?

Spark X2.5 4B needs 2.3 GB for its weights at 4-bit, 3.4 GB with an 8K context and 4.3 GB with 32K. At 8-bit, 5.2 GB.

What is the minimum VRAM for Spark X2.5 4B?

About 3.0 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 Spark X2.5 4B?

4.1B parameters: 7.7 GB of weights at 16-bit, 2.3 GB at 4-bit.

Can Spark X2.5 4B run on an RTX 4090?

Yes. At 8-bit it holds 480K of context and writes about 138 tokens a second.

More models