Skip to content

Diffusiongemma 26B A4B IT VRAM requirements

How much VRAM Diffusiongemma 26B A4B IT needs
4-bit weights: 14.6 GBWith 8K context: 16.0 GBWith 32K context: 16.4 GB8-bit, 8K context: 26.9 GB
Parameters
25.8B
Active per token
4B
Longest context
256K
Released
Jun 2026
License
Apache 2.0
Strongest at
Text and images

How much VRAM Diffusiongemma 26B A4B IT needs

Memory Diffusiongemma 26B A4B IT needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit48.1 GB49.5 GB50.0 GB51.8 GB
8-bit25.6 GB26.9 GB27.4 GB29.3 GB
6-bit19.7 GB21.1 GB21.6 GB23.5 GB
5-bit17.1 GB18.5 GB19.0 GB20.8 GB
4-bit14.6 GB16.0 GB16.4 GB18.3 GB
3-bit11.8 GB13.1 GB13.6 GB15.5 GB
2-bit9.0 GB10.4 GB10.9 GB12.8 GB
1.6-bit5.0 GB6.3 GB6.8 GB8.7 GB

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

GPUs and Macs that run Diffusiongemma 26B A4B IT

The best way Diffusiongemma 26B A4B IT runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Diffusiongemma 26B A4B IT on RTX 4060 report
RTX 5060Too bigPDF: Diffusiongemma 26B A4B IT on RTX 5060 report
RTX 3070Too bigPDF: Diffusiongemma 26B A4B IT on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Diffusiongemma 26B A4B IT on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: Diffusiongemma 26B A4B IT on RTX 3060 12 GB report
Arc B580Too bigPDF: Diffusiongemma 26B A4B IT on Arc B580 report
RTX 4070Too bigPDF: Diffusiongemma 26B A4B IT on RTX 4070 report
RTX 4070 SuperToo bigPDF: Diffusiongemma 26B A4B IT on RTX 4070 Super report
RTX 5070Too bigPDF: Diffusiongemma 26B A4B IT on RTX 5070 report
RTX 4060 Ti 16 GB3-bit112K87 tok/s1.1 sPDF: Diffusiongemma 26B A4B IT on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB3-bit112K124 tok/s1.0 sPDF: Diffusiongemma 26B A4B IT on RTX 5060 Ti 16 GB report
Arc A770 16 GB3-bit112K100 tok/s1.2 sPDF: Diffusiongemma 26B A4B IT on Arc A770 16 GB report
Radeon RX 9070 XT3-bit112K128 tok/s0.7 sPDF: Diffusiongemma 26B A4B IT on Radeon RX 9070 XT report
RTX 4070 Ti Super3-bit112K168 tok/s0.6 sPDF: Diffusiongemma 26B A4B IT on RTX 4070 Ti Super report
RTX 4080 Super3-bit112K179 tok/s0.5 sPDF: Diffusiongemma 26B A4B IT on RTX 4080 Super report
RTX 5070 Ti3-bit112K203 tok/s0.6 sPDF: Diffusiongemma 26B A4B IT on RTX 5070 Ti report
RTX 50803-bit112K212 tok/s0.4 sPDF: Diffusiongemma 26B A4B IT on RTX 5080 report
Radeon RX 7900 XTX6-bit88K119 tok/s1.1 sPDF: Diffusiongemma 26B A4B IT on Radeon RX 7900 XTX report
RTX 30906-bit88K147 tok/s0.7 sPDF: Diffusiongemma 26B A4B IT on RTX 3090 report
RTX 40906-bit88K155 tok/s0.3 sPDF: Diffusiongemma 26B A4B IT on RTX 4090 report
Radeon AI PRO R97008-bit184K73 tok/s0.7 sPDF: Diffusiongemma 26B A4B IT on Radeon AI PRO R9700 report
RTX 50908-bit184K193 tok/s0.2 sPDF: Diffusiongemma 26B A4B IT on RTX 5090 report
2× RTX 30908-bit256K121 tok/s0.4 sPDF: Diffusiongemma 26B A4B IT on 2× RTX 3090 report
2× RTX 40908-bit256K128 tok/s0.2 sPDF: Diffusiongemma 26B A4B IT on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit256K158 tok/s0.4 sPDF: Diffusiongemma 26B A4B IT on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit256K123 tok/s0.3 sPDF: Diffusiongemma 26B A4B IT on RTX 6000 Ada report
L40S8-bit256K113 tok/s0.4 sPDF: Diffusiongemma 26B A4B IT on L40S report
2× RTX 50908-bit256K193 tok/s0.1 sPDF: Diffusiongemma 26B A4B IT on 2× RTX 5090 report
4× RTX 30908-bit256K121 tok/s0.2 sPDF: Diffusiongemma 26B A4B IT on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit256K193 tok/s0.2 sPDF: Diffusiongemma 26B A4B IT on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Diffusiongemma 26B A4B IT on Mac mini M4 16 GB report
Mac mini M4 24 GB3-bit120K42 tok/s16 sPDF: Diffusiongemma 26B A4B IT on Mac mini M4 24 GB report
Mac mini M4 32 GB5-bit104K30 tok/s16 sPDF: Diffusiongemma 26B A4B IT on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB3-bit120K77 tok/s7.8 sPDF: Diffusiongemma 26B A4B IT on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit256K43 tok/s7.8 sPDF: Diffusiongemma 26B A4B IT on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit256K43 tok/s7.8 sPDF: Diffusiongemma 26B A4B IT on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit256K69 tok/s3.9 sPDF: Diffusiongemma 26B A4B IT on M4 Max 128 GB report
M5 Max 128 GB8-bit256K74 tok/s1.1 sPDF: Diffusiongemma 26B A4B IT on M5 Max 128 GB report
M3 Ultra 96 GB8-bit256K82 tok/s2.2 sPDF: Diffusiongemma 26B A4B IT on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit256K82 tok/s2.2 sPDF: Diffusiongemma 26B A4B IT on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit256K82 tok/s2.2 sPDF: Diffusiongemma 26B A4B IT on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit256K104 tok/s0.7 sPDF: Diffusiongemma 26B A4B IT on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit256K33 tok/s2.3 sPDF: Diffusiongemma 26B A4B IT on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit256K33 tok/s2.3 sPDF: Diffusiongemma 26B A4B IT on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit256K45 tok/s0.8 sPDF: Diffusiongemma 26B A4B IT on RTX Spark report
DGX Spark8-bit256K42 tok/s0.8 sPDF: Diffusiongemma 26B A4B IT on DGX Spark report
Dual DGX Spark8-bit256K42 tok/s0.5 sPDF: Diffusiongemma 26B A4B IT on Dual DGX Spark report
DGX Station GB3008-bit256K377 tok/s0.1 sPDF: Diffusiongemma 26B A4B IT on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit256K193 tok/s0.1 sPDF: Diffusiongemma 26B A4B IT on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit256K193 tok/s0.1 sPDF: Diffusiongemma 26B A4B IT on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit256K193 tok/s0.1 sPDF: Diffusiongemma 26B A4B IT on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit256K209 tok/s0.5 sPDF: Diffusiongemma 26B A4B IT on A100 80 GB SXM report
H100 SXM8-bit256K277 tok/s0.2 sPDF: Diffusiongemma 26B A4B IT on H100 SXM report
4× H200 NVL8-bit256K326 tok/s0.1 sPDF: Diffusiongemma 26B A4B IT on 4× H200 NVL report
8× H200 NVL8-bit256K326 tok/s0.1 sPDF: Diffusiongemma 26B A4B IT 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 unsloth/diffusiongemma-26B-A4B-it-GGUF
vLLM
vllm serve google/diffusiongemma-26B-A4B-it
MLX (Mac)
mlx_lm.chat --model mlx-community/diffusiongemma-26B-A4B-it-4bit

Questions

How much VRAM does Diffusiongemma 26B A4B IT need?

Diffusiongemma 26B A4B IT needs 14.6 GB for its weights at 4-bit, 16.0 GB with an 8K context and 16.4 GB with 32K. At 8-bit, 26.9 GB.

What is the minimum VRAM for Diffusiongemma 26B A4B IT?

About 13.1 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 Ti 16 GB.

How big is Diffusiongemma 26B A4B IT?

25.8B parameters (4B active per token): 48.1 GB of weights at 16-bit, 14.6 GB at 4-bit.

Can Diffusiongemma 26B A4B IT run on an RTX 4090?

Yes. At 6-bit it holds 88K of context and writes about 155 tokens a second.

More models