Gemma 4 26B-A4B hardware requirements

What is Gemma 4 26B-A4B?

Gemma 4 26B-A4B is the mixture-of-experts model in Google DeepMind's Gemma 4 family, released in April 2026. It has about 26B parameters and activates about 4B per token, but all the weights still have to fit in VRAM or system RAM. It reads text and images with up to a 256k context, targets reasoning, coding and function calling, keeps thinking off by default and is licensed under Apache 2.0.

Model card · Release post

What to know before running Gemma 4 26B-A4B

Gemma 4 26B-A4B is Google's mixture-of-experts Gemma: about 26B parameters in total, about 4B used per token. At Q4_K_M it needs 17.7 GB with an 8k context, so a 24GB card holds it completely and runs it great. A 16GB card is where the quant matters. The IQ4_XS file, at 13.6 GB, fits fully on the GeForce RTX 5060 Ti 16GB and runs great, while the Q4_K_M file keeps some experts in system RAM and runs well, by a theoretical estimate (±30%). Our quantization guide covers what IQ4_XS gives up.

On a 12GB card such as the GeForce RTX 3060 12GB it runs well with its experts partly in system RAM; on the 8GB GeForce RTX 4060 it only just clears the bar and runs well too. Those are theoretical estimates (±30%). Its thinking mode is off by default, but CanRun still holds models that have one to a 1.5× stricter speed bar, so with thinking left off it may do a grade better than the table shows. Its KV cache is small, so long contexts cost little memory, and it reads images as well as text. On a Mac, the 48GB Apple M4 Pro runs it great.

About the model

Parameters: 26.54B total · 4B active (MoE)
Attention: Sliding-window attention
Max context: 262,144 tokens
Released: 2026-03-11
Model card: Gemma 4 26B-A4B on Hugging Face

Other sizes in this series

Graphics card memory each model needs at its baseline quant with 8k context, and its verdict on common graphics cards, smallest model first
Model	Parameters	Baseline quant	Memory needed (8k)	GeForce RTX 4060	GeForce RTX 5060 Ti 16GB	GeForce RTX 3090
Gemma 4 E2B	5.12B	Q4_K_M	2.1 GB	Runs great	Runs great	Runs great
Gemma 4 E4B	8B	Q4_K_M	3.7 GB	Runs great	Runs great	Runs great
Gemma 4 12B	11.96B	Q4_K_M	7.8 GB	Runs well	Runs great	Runs great
Gemma 4 26B-A4Bthis model	26.54B total · 4B active (MoE)	Q4_K_M	17.7 GB	Runs well	Runs well	Runs great
Gemma 4 31B	31.27B	Q4_K_M	19.5 GB	Runs slowly	Runs slowly	Runs great

GGUF files

Tracked quantizations of Gemma 4 26B-A4B, smallest first
Quant	File size	Quality	Published by
UD-Q2_K_XL	10.55 GB	Heavy loss	Community GGUF · unsloth
IQ4_XS	13.60 GB	Good balance	Community GGUF · unsloth
Q4_K_Mbaseline	16.95 GB	Good balance	Community GGUF · unsloth
Q8_0	26.86 GB	Near-lossless	Community GGUF · unsloth

File sizes do not include the vision projector (mmproj) that image input needs.

Some repositories publish only Unsloth Dynamic (UD) variants; sizes are those of the file you would download.

Memory needed by quant and context

Graphics card memory in GB: weights loaded on the GPU + f16 KV cache + compute buffer. Add about 0.6 GB if the GPU also drives your display on Windows.
Quant	4k	8k	16k	32k	64k	128k	256k
UD-Q2_K_XL	11.2	11.3	11.5	12.0	13.0	14.9	18.8
IQ4_XS	14.2	14.3	14.6	15.1	16.0	18.0	21.8
Q4_K_M	17.6	17.7	17.9	18.4	19.4	21.3	25.2
Q8_0	27.5	27.6	27.8	28.3	29.3	31.2	35.1

Which GPUs can run Gemma 4 26B-A4B?

How to read the verdicts

Runs great: Fully on the GPU at 20 tok/s or more
Runs well: Fully on the GPU at 8–20 tok/s; MoE experts in system RAM at 20 tok/s or more; or at least 90% on the GPU at 8 tok/s or more
Runs slowly: 2–8 tok/s; CPU-only; less than 90% on the GPU; or MoE experts in system RAM below 20 tok/s
Won't run: Does not fit, or under 2 tok/s

Sort

Filter this list by budget, type and context in the calculator

Scroll sideways to see every column.

58 GPUs, Macs and CPU setups at Q4_K_M with 8k context, sorted by verdict and speed
Hardware	Verdict	Speed	Quant	Memory	Runs as	Context	Notes	Price
GeForce RTX 5090	Runs greatGemma 4 26B-A4B verdict on the GeForce RTX 5090	216.1 tok/s8k estimate 210.4 tok/s (168.3–252.4, calibrated estimate ±20%)measured (2 runs, 512/8k context)	Q4_K_M	18.3 / 32.0 GB	Full GPU	up to 256k	—	$4,200 street (as of 2026-08-09)
NVIDIA RTX PRO 6000 Blackwell	Runs great	est. 210.4 tok/s168.3–252.4calibrated estimate ±20%	Q4_K_M	18.3 / 96.0 GB	Full GPU	up to 256k	—	—
NVIDIA RTX PRO 6000 Blackwell Max-Q	Runs great	est. 210.4 tok/s168.3–252.4calibrated estimate ±20%	Q4_K_M	18.3 / 96.0 GB	Full GPU	up to 256k	—	—
GeForce RTX 3090 Ti	Runs great	est. 142.6 tok/s114.1–171.1calibrated estimate ±20%	Q4_K_M	18.3 / 24.0 GB	Full GPU	up to 128k	—	$1,999 launch MSRP
GeForce RTX 4090	Runs greatGemma 4 26B-A4B verdict on the GeForce RTX 4090	est. 142.6 tok/s114.1–171.1calibrated estimate ±20%	Q4_K_M	18.3 / 24.0 GB	Full GPU	up to 128k	—	$2,755 street (as of 2026-08-09)
GeForce RTX 5090 Laptop	Runs great	est. 130.6 tok/s104.5–156.7calibrated estimate ±20%	Q4_K_M	18.3 / 24.0 GB	Full GPU	up to 128k	—	—
Apple M5 Ultra · 96 GB	Runs great	est. 115.8 tok/s92.6–138.9calibrated estimate ±20%	Q4_K_M	17.7 / 67.2 GB	Unified memory	up to 256k	—	—
Radeon RX 7900 XTX	Runs great	est. 114.7 tok/s91.8–137.7calibrated estimate ±20%	Q4_K_M	18.3 / 24.0 GB	Full GPU	up to 128k	—	$999 launch MSRP
GeForce RTX 3090	Runs greatGemma 4 26B-A4B verdict on the GeForce RTX 3090	108.0 tok/s8k estimate 134.9 tok/s (108.0–161.9, calibrated estimate ±20%)measured (1 run, 32k context)	Q4_K_M	18.3 / 24.0 GB	Full GPU	up to 128k	—	$1,050 street (as of 2026-08-09)
Radeon RX 7900 XT	Runs great	est. 102.9 tok/s82.4–123.5calibrated estimate ±20%	Q4_K_M	18.3 / 20.0 GB	Full GPU	up to 64k	—	$899 launch MSRP
Apple M5 Max (40-core GPU) · 48 GB	Runs great	est. 98.4 tok/s78.7–118.1calibrated estimate ±20%	Q4_K_M	17.7 / 33.6 GB	Unified memory	up to 256k	—	—
Apple M4 Max (40-core GPU) · 48 GB	Runs great	est. 89.8 tok/s71.8–107.8calibrated estimate ±20%	Q4_K_M	17.7 / 33.6 GB	Unified memory	up to 256k	—	—
Apple M3 Ultra · 96 GB	Runs great	est. 89.0 tok/s71.2–106.8calibrated estimate ±20%	Q4_K_M	17.7 / 67.2 GB	Unified memory	up to 256k	—	—
Apple M2 Ultra · 64 GB	Runs great	79.1 tok/s8k estimate 87.5 tok/s (70.0–105.0, calibrated estimate ±20%)measured (1 run, 512 context)	Q4_K_M	17.7 / 44.8 GB	Unified memory	up to 256k	—	—
Apple M5 Max (32-core GPU) · 36 GB	Runs great	est. 78.3 tok/s62.6–93.9calibrated estimate ±20%	Q4_K_M	17.7 / 25.2 GB	Unified memory	up to 128k	—	—
Apple M4 Max (32-core GPU) · 36 GB	Runs great	est. 71.2 tok/s56.9–85.4calibrated estimate ±20%	Q4_K_M	17.7 / 25.2 GB	Unified memory	up to 128k	—	—
Apple M5 Pro · 48 GB	Runs great	est. 55.6 tok/s44.5–66.8calibrated estimate ±20%	Q4_K_M	17.7 / 33.6 GB	Unified memory	up to 128k	—	—
Ryzen AI Max+ 395 (Strix Halo) · 32 GB	Runs great	54.3 tok/s8k estimate 48.1 tok/s (38.5–57.7, calibrated estimate ±20%)measured (2 runs, 512 context)	Q4_K_M	17.7 / 22.4 GB	Unified memory	up to 128k	—	$1,999 launch MSRP (128 GB)
Apple M4 Pro · 48 GB	Runs greatGemma 4 26B-A4B verdict on the Apple M4 Pro	est. 50.2 tok/s40.2–60.2calibrated estimate ±20%	Q4_K_M	17.7 / 33.6 GB	Unified memory	up to 128k	—	—
NVIDIA DGX Spark · 128 GB	Runs great	47.0 tok/s8k estimate 51.5 tok/s (41.2–61.8, calibrated estimate ±20%)measured (1 run, 32k context)	Q4_K_M	17.7 / 116.9 GB	Unified memory	up to 128k	—	$3,999 launch MSRP (128 GB)
Apple M6 · 32 GB	Runs great	est. 32.8 tok/s26.3–39.4calibrated estimate ±20%	Q4_K_M	17.7 / 22.4 GB	Unified memory	up to 16k	—	—
GeForce RTX 5080	Runs well	est. 90.8 tok/s63.6–118.1theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 256k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	$1,256 street (as of 2026-08-09)
GeForce RTX 5070 Ti	Runs well	est. 87.9 tok/s61.5–114.3theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 256k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	$949 street (as of 2026-08-09)
GeForce RTX 5080 Laptop	Runs well	est. 87.9 tok/s61.5–114.3theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 256k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	—
GeForce RTX 4080 Super	Runs well	est. 79.5 tok/s55.7–103.4theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 256k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	$999 launch MSRP
GeForce RTX 4080	Runs well	est. 78.4 tok/s54.9–101.9theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 256k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	$1,199 launch MSRP
GeForce RTX 4070 Ti Super	Runs well	est. 75.7 tok/s53.0–98.3theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	$799 launch MSRP
GeForce RTX 4090 Laptop	Runs well	est. 69.2 tok/s48.5–90.0theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	—
Radeon RX 9070	Runs well	est. 67.6 tok/s47.3–87.9theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	$639 street (as of 2026-08-09)
Radeon RX 9070 XT	Runs well	est. 67.6 tok/s47.3–87.9theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	$689 street (as of 2026-08-09)
Radeon RX 7800 XT	Runs well	est. 66.5 tok/s46.6–86.5theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	$499 launch MSRP
GeForce RTX 3080 12GB	Runs well	est. 59.3 tok/s41.5–77.1theoretical estimate ±30%	Q4_K_M	11.1 / 12.0 GB + 7.2 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory UD-Q2_K_XL (heavy quality loss): Runs great	$799 launch MSRP
GeForce RTX 5060 Ti 16GB	Runs wellGemma 4 26B-A4B verdict on the GeForce RTX 5060 Ti 16GB	est. 59.2 tok/s41.4–76.9theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	$569 street (as of 2026-08-09)
GeForce RTX 5070	Runs wellGemma 4 26B-A4B verdict on the GeForce RTX 5070	est. 53.6 tok/s37.5–69.7theoretical estimate ±30%	Q4_K_M	11.1 / 12.0 GB + 7.2 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory UD-Q2_K_XL (heavy quality loss): Runs great	$629 street (as of 2026-08-09)
GeForce RTX 5070 Ti Laptop	Runs well	est. 53.6 tok/s37.5–69.7theoretical estimate ±30%	Q4_K_M	11.1 / 12.0 GB + 7.2 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory UD-Q2_K_XL (heavy quality loss): Runs great	—
GeForce RTX 3080 10GB	Runs well	est. 49.2 tok/s34.5–64.0theoretical estimate ±30%	Q4_K_M	9.2 / 10.0 GB + 9.1 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$699 launch MSRP
GeForce RTX 2080 Ti	Runs well	est. 48.9 tok/s34.2–63.5theoretical estimate ±30%	Q4_K_M	10.1 / 11.0 GB + 8.2 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$999 launch MSRP
GeForce RTX 4070	Runs wellGemma 4 26B-A4B verdict on the GeForce RTX 4070	est. 47.8 tok/s33.5–62.1theoretical estimate ±30%	Q4_K_M	11.1 / 12.0 GB + 7.2 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory UD-Q2_K_XL (heavy quality loss): Runs great	$599 launch MSRP
GeForce RTX 4070 Super	Runs well	est. 47.8 tok/s33.5–62.1theoretical estimate ±30%	Q4_K_M	11.1 / 12.0 GB + 7.2 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory UD-Q2_K_XL (heavy quality loss): Runs great	$599 launch MSRP
Radeon RX 9060 XT 16GB	Runs well	est. 45.7 tok/s32.0–59.4theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	$449 street (as of 2026-08-09)
GeForce RTX 4080 Laptop	Runs well	est. 44.6 tok/s31.2–58.0theoretical estimate ±30%	Q4_K_M	11.1 / 12.0 GB + 7.2 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory UD-Q2_K_XL (heavy quality loss): Runs great	—
GeForce GTX 1080 Ti	Runs well	est. 44.5 tok/s31.2–57.9theoretical estimate ±30%	Q4_K_M	10.1 / 11.0 GB + 8.2 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$699 launch MSRP
GeForce RTX 4060 Ti 16GB	Runs well	est. 43.4 tok/s30.4–56.4theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	$499 launch MSRP
GeForce RTX 3060 12GB	Runs wellGemma 4 26B-A4B verdict on the GeForce RTX 3060 12GB	est. 40.7 tok/s28.5–53.0theoretical estimate ±30%	Q4_K_M	11.1 / 12.0 GB + 7.2 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory UD-Q2_K_XL (heavy quality loss): Runs great	$250 street (as of 2026-08-09)
Intel Arc B580	Runs well	est. 39.0 tok/s27.3–50.8theoretical estimate ±30%	Q4_K_M	11.1 / 12.0 GB + 7.2 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory UD-Q2_K_XL (heavy quality loss): Runs great	$290 street (as of 2026-08-09)
GeForce RTX 3070	Runs well	est. 37.6 tok/s26.3–48.8theoretical estimate ±30%	Q4_K_M	7.3 / 8.0 GB + 11.0 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$499 launch MSRP
GeForce RTX 5060	Runs well	est. 37.6 tok/s26.3–48.8theoretical estimate ±30%	Q4_K_M	7.3 / 8.0 GB + 11.0 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$339 street (as of 2026-08-09)
GeForce RTX 5060 Ti 8GB	Runs well	est. 37.6 tok/s26.3–48.8theoretical estimate ±30%	Q4_K_M	7.3 / 8.0 GB + 11.0 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$429 street (as of 2026-08-09)
GeForce RTX 4060 Ti 8GB	Runs well	est. 31.4 tok/s22.0–40.8theoretical estimate ±30%	Q4_K_M	7.3 / 8.0 GB + 11.0 GB RAM	MoE experts in RAM	up to 16k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$399 launch MSRP
GeForce RTX 4060	Runs wellGemma 4 26B-A4B verdict on the GeForce RTX 4060	est. 30.6 tok/s21.4–39.7theoretical estimate ±30%	Q4_K_M	7.3 / 8.0 GB + 11.0 GB RAM	MoE experts in RAM	up to 8k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$299 launch MSRP
Apple M5 · 32 GB	Runs well	est. 29.8 tok/s23.8–35.7calibrated estimate ±20%	Q4_K_M	17.7 / 22.4 GB	Unified memory	up to 128k	Try IQ4_XS: Runs great	—
Apple M4 · 32 GB	Runs wellGemma 4 26B-A4B verdict on the Apple M4	est. 23.6 tok/s18.9–28.4calibrated estimate ±20%	Q4_K_M	17.7 / 22.4 GB	Unified memory	up to 128k	UD-Q2_K_XL (heavy quality loss): Runs great	—
GeForce RTX 4060 Laptop	Runs slowly	est. 29.7 tok/s20.8–38.6theoretical estimate ±30%	Q4_K_M	7.3 / 8.0 GB + 11.0 GB RAM	MoE experts in RAM	up to 128k	Experts run from system RAM — Runs well needs 30 tok/s or more on this path Try IQ4_XS: Runs well	—
GeForce RTX 4070 Laptop	Runs slowly	est. 29.7 tok/s20.8–38.6theoretical estimate ±30%	Q4_K_M	7.3 / 8.0 GB + 11.0 GB RAM	MoE experts in RAM	up to 128k	Experts run from system RAM — Runs well needs 30 tok/s or more on this path Try IQ4_XS: Runs well	—
GeForce RTX 3050 8GB	Runs slowly	est. 27.7 tok/s19.4–36.1theoretical estimate ±30%	Q4_K_M	7.3 / 8.0 GB + 11.0 GB RAM	MoE experts in RAM	up to 128k	Experts run from system RAM — Runs well needs 30 tok/s or more on this path Try IQ4_XS: Runs well	$249 launch MSRP
DDR5-6000 dual-channel CPU · 32 GB	Runs slowly	est. 15.2 tok/s12.1–18.2calibrated estimate ±20%	Q4_K_M	17.1 GB RAM	CPU only	up to 256k	CPU inference — capped at Runs slowly	—
DDR5-5600 dual-channel CPU · 32 GB	Runs slowly	est. 14.2 tok/s11.3–17.0calibrated estimate ±20%	Q4_K_M	17.1 GB RAM	CPU only	up to 256k	CPU inference — capped at Runs slowly	—
DDR4-3200 dual-channel CPU · 32 GB	Runs slowly	est. 8.1 tok/s6.5–9.7calibrated estimate ±20%	Q4_K_M	17.1 GB RAM	CPU only	up to 128k	—	—

Models with a thinking mode spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s). With thinking turned off, a model may do one grade better.

Cheapest GPUs that run it

Runs greatGeForce RTX 3090$1,050 street (as of 2026-08-09), 108.0 tok/s measured at 32k context (1 run; 8k estimate 134.9 tok/s, 108.0–161.9, calibrated estimate ±20%).
Runs wellGeForce RTX 3060 12GB$250 street (as of 2026-08-09), est. 40.7 tok/s (28.5–53.0, theoretical estimate ±30%).

Measured results for Gemma 4 26B-A4B

Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.

Measured results for Gemma 4 26B-A4B
Hardware	Quant	Backend	Context	Prompt (tok/s)	Generation (tok/s)	Flags	Source	Measured
GeForce RTX 5090	Q4_K_M	llama.cpp	512	13,588.0	223.8	CUDA, KV q8_0	github.com	2026-05-24
GeForce RTX 5090	Q4_K_M	ollama	8k	—	208.4	CUDA, FA off	github.com	2026-04-05
GeForce RTX 3090	Q4_K_M	ollama	32k	—	108.0	FA off	github.com	2026-04-07
Radeon RX 7900 XT	UD-Q4_K_XL	llama.cpp	512	2,884.0	77.8	ROCm	github.com	2026-04-18
Apple M2 Ultra	Q8_0	llama.cpp	512	2,160.0	87.4	Metal	github.com	2026-09-15
Apple M5 Max (40-core GPU)	Q8_0	llama.cpp	512	3,582.0	95.4	Metal	github.com	2026-09-15
Apple M2 Ultra	Q4_K_M	llama.cpp	512	1,550.0	79.1	Metal	github.com	2026-09-19
Ryzen AI Max+ 395 (Strix Halo)	Q4_K_M	llama.cpp	512	1,327.0	55.5	Vulkan	github.com	2026-06-30
Ryzen AI Max+ 395 (Strix Halo)	Q4_K_M	llama.cpp	512	1,209.0	53.0	Vulkan	github.com	2026-08-18
Ryzen AI Max+ 395 (Strix Halo)	UD-Q4_K_XL	llama.cpp	512	1,325.0	54.7	Vulkan	github.com	2026-05-18
NVIDIA DGX Spark	Q4_K_M	llama.cpp	32k	—	47.0	CUDA, depth 32768	github.com	2026-06-11

Frequently asked questions

How much VRAM does Gemma 4 26B-A4B need?

At Q4_K_M the file is 16.95 GB; with 8k context a graphics card needs about 17.7 GB (weights 16.93 GB, KV cache 0.2 GB, compute buffer 0.6 GB). The smallest tracked file (UD-Q2_K_XL (heavy quality loss)) needs about 11.3 GB.

What is the cheapest GPU that runs Gemma 4 26B-A4B well?

Cheapest that runs it great: GeForce RTX 3090 at $1,050 (street price as of 2026-08-09), 108.0 tok/s measured at 32k context (1 run; 8k estimate 134.9 tok/s, 108.0–161.9, calibrated estimate ±20%). Cheaper and still Runs well: GeForce RTX 3060 12GB at $250 (street price as of 2026-08-09), est. 40.7 tok/s (28.5–53.0, theoretical estimate ±30%).

Can Gemma 4 26B-A4B run on an 8 GB, 16 GB or 24 GB GPU?

At Q4_K_M with 8k context: GeForce RTX 4060 — Runs well, est. 30.6 tok/s (21.4–39.7, theoretical estimate ±30%); GeForce RTX 5060 Ti 16GB — Runs well, est. 59.2 tok/s (41.4–76.9, theoretical estimate ±30%); GeForce RTX 3090 — Runs great, 108.0 tok/s measured at 32k context (1 run; 8k estimate 134.9 tok/s, 108.0–161.9, calibrated estimate ±20%).

Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.