How much memory does Gemma-4-26B-A4B at BF16 need?

52GB. That is 50GB of weights at 16 bits per weight, 0.3GB of K/V cache at 8K context, and 1.2GB of runtime overhead.

Which machine is cheapest for Gemma-4-26B-A4B at BF16?

Mac mini, M5 Pro, 64GB, $2,899 at Apple. It fits with 7GB of headroom and decodes at an est. 20.2 tok/s.

Cheapest box to run Gemma-4-26B-A4B at BF16

Google’s Gemma-4-26B-A4B is 25.2B parameters, 3.8B of them read per token. The number below is computed from that and each vendor’s published specs, not asserted. Switch the quant or the context and it recomputes in place.

Fit finderprices checked 2 hours ago

Quant

Context

Gemma-4-26B-A4B at BF16 needs 52GB.

Cheapest Config that fits: Mac mini, M5 Pro, 64GB. Fits, 7GB headroom. Est. 20.2 tok/s.

Best offer today: $2,899 at Apple — in stock, checked 2 hours ago.

Price history: 34 days recorded.

The 26B in the name is marketing; Google's own table reads 25.2B total and 3.8B active, and that is what we publish. Only the 5 global-attention layers cache unboundedly; the rest slide a 1024-token window.

Offers — Mac mini, M5 Pro, 64GB, 512GB

B&H PhotoPreorder, ships about 22 Sep. Below Apple's own price for the same build, which is unusual enough that we are re-checking both numbers.out of stockchecked 33 days ago$2,699
Applelist pricein stockchecked 2 hours ago$2,899

Mac mini, M6, 16GB$1,099
Won't fit−39GB headroom
Mac mini, M6, 24GB$1,299
Won't fit−31GB headroom
Mac mini, M6, 32GB$1,499
Won't fit−23GB headroom
Mac mini, M5 Pro, 24GB$1,699
Won't fit−31GB headroom
Mac mini, M5 Pro, 48GB$2,499
Won't fit−8GB headroom
Mac Studio, M5 Max, 36GB$2,499
Won't fit−19GB headroom
Mac mini, M5 Pro, 64GB$2,899
Fits7GB headroomest. 20.2 tok/s
GeForce RTX 5090, 32GB$2,999.99+host
Won't fit−21GB headroom
Mac Studio, M4 Max, 64GB, refurbished$3,049
Fits7GB headroomest. 35.9 tok/s
Mac Studio, M5 Max, 48GB$3,099
Won't fit−8GB headroom
Mac Studio, M5 Max, 64GB$3,499
Fits7GB headroomest. 40.4 tok/s
Ascent GX10, GB10 Grace Blackwell, 128GB$3,999.99
Fits68GB headroomest. 18.0 tok/s
Mac Studio, M3 Ultra, 96GB, refurbished$4,329
Fits36GB headroomest. 53.9 tok/s
Mac Studio, M4 Max, 128GB, refurbished$4,409
Fits68GB headroomest. 35.9 tok/s
DGX Spark, GB10 Grace Blackwell, 128GB$4,699
Fits68GB headroomest. 18.0 tok/s
Mac Studio, M5 Max, 128GB$5,099
Fits68GB headroomest. 40.4 tok/s
Mac Studio, M5 Ultra, 96GB$5,499
Fits36GB headroomest. 78.9 tok/s
Mac Studio, M3 Ultra, 256GB, refurbished$7,729
Fits196GB headroomest. 53.9 tok/s
Mac Studio, M5 Ultra, 256GB$9,499
Fits196GB headroomest. 78.9 tok/s
Mac Studio, M5 Ultra, 512GB—
Fits452GB headroomest. 78.9 tok/s

Config	Memory	Price	$/GB	Fit	Headroom	Est. speed
Mac mini, M6	16GB	$1,099	$69	Won't fit	−39GB	—
Mac mini, M6	24GB	$1,299	$54	Won't fit	−31GB	—
Mac mini, M6	32GB	$1,499	$47	Won't fit	−23GB	—
Mac mini, M5 Pro	24GB	$1,699	$71	Won't fit	−31GB	—
Mac mini, M5 Pro	48GB	$2,499	$52	Won't fit	−8GB	—
Mac Studio, M5 Max	36GB	$2,499	$69	Won't fit	−19GB	—
Mac mini, M5 Pro	64GB	$2,899	$45	Fits	7GB	20.2 tok/s
GeForce RTX 5090	32GB	$2,999.99+host	$94	Won't fit	−21GB	—
Mac Studio, M4 Max, refurb	64GB	$3,049	$48	Fits	7GB	35.9 tok/s
Mac Studio, M5 Max	48GB	$3,099	$65	Won't fit	−8GB	—
Mac Studio, M5 Max	64GB	$3,499	$55	Fits	7GB	40.4 tok/s
Ascent GX10, GB10 Grace Blackwell	128GB	$3,999.99	$31	Fits	68GB	18.0 tok/s
Mac Studio, M3 Ultra, refurb	96GB	$4,329	$45	Fits	36GB	53.9 tok/s
Mac Studio, M4 Max, refurb	128GB	$4,409	$34	Fits	68GB	35.9 tok/s
DGX Spark, GB10 Grace Blackwell	128GB	$4,699	$37	Fits	68GB	18.0 tok/s
Mac Studio, M5 Max	128GB	$5,099	$40	Fits	68GB	40.4 tok/s
Mac Studio, M5 Ultra	96GB	$5,499	$57	Fits	36GB	78.9 tok/s
Mac Studio, M3 Ultra, refurb	256GB	$7,729	$30	Fits	196GB	53.9 tok/s
Mac Studio, M5 Ultra	256GB	$9,499	$37	Fits	196GB	78.9 tok/s
Mac Studio, M5 Ultra	512GB	—	—	Fits	452GB	78.9 tok/s

GeForce RTX 5090 is a card, not a computer: every price shown for it is the cheapest listing we found plus $1,000 for the minimum host machine.

memory = 25.2B × 2.000 B/param (16 bits) + 0.042 GB/1K × 8K K/V + 1.2 GB runtime
speed = bandwidth × 0.5 ÷ (3.8B active × 2.000 B/param) = 7.6 GB read per token
usable = memory − operating system reserve (8% on a Mac, 6% on a DGX Spark, 4% on a card)

Bandwidth is the vendor’s published figure. Every speed is an estimate and says so. Methodology.

What else Mac mini M5 Pro 64GB runs

Llama-4-Scout-17B-16EQ3_K_M, est. 18.5 tok/s
Qwen3-Coder-NextQ5_K_M, est. 72.2 tok/s
Llama-3.3-70BQ5_K_M, est. 4.0 tok/s
Qwen3.5-35B-A3BQ8_0, est. 48.2 tok/s
Qwen3-32BQ8_0, est. 5.9 tok/s
Gemma-4-31BQ8_0, est. 6.1 tok/s

Best quant that still fits at 8K context, largest model first. Full list on the Config page.

Bandwidth 307GB/s, usable memory 59GB of 64GB.