DiffusionGemma 26B A4B

Google · google/diffusiongemma-26b-a4b-it

Provider	Google
Family	gemma
Type	image-generation
Status	active
Released	2026-06-09
Updated	2026-07-15
Parameters	25.8B
Open weights	yes

What it fits on

Ordered by predicted speed. Bandwidth sets decode rate; memory decides whether it runs at all.

Every figure here is computed, not measured. Fit is weights at each quantisation against device memory, with a 25% allowance for the KV cache, activations and the OS. Decode rate is the memory-bandwidth roofline at 70% efficiency. Nobody has run this model on these devices. This is a mixture-of-experts model and no card carries active-parameter counts, so these predictions use total parameters and understate the real speed.

Device	Memory	Bandwidth	Best quality	Weights	tok/s	Fastest
Cerebras WSE-3	44.0 GB	21000000.0 GB/s	fp8	25.82 GB	~569242.8	~1138485.6 q4
NVIDIA GB200 Grace Blackwell Superchip	372.0 GB	16000.0 GB/s	bf16	51.65 GB	~216.9	~867.4 q4
AMD Instinct MI355X	288.0 GB	8000.0 GB/s	bf16	51.65 GB	~108.4	~433.7 q4
NVIDIA B200	180.0 GB	7700.0 GB/s	bf16	51.65 GB	~104.4	~417.4 q4
Google TPU7x (Ironwood)	192.0 GB	7380.0 GB/s	bf16	51.65 GB	~100.0	~400.1 q4
AMD Instinct MI325X	256.0 GB	6000.0 GB/s	bf16	51.65 GB	~81.3	~325.3 q4
AMD Instinct MI300X	192.0 GB	5300.0 GB/s	bf16	51.65 GB	~71.8	~287.3 q4
NVIDIA H200 SXM	141.0 GB	4800.0 GB/s	bf16	51.65 GB	~65.1	~260.2 q4
NVIDIA H100 NVL	94.0 GB	3900.0 GB/s	bf16	51.65 GB	~52.9	~211.4 q4
NVIDIA H100 SXM	80.0 GB	3350.0 GB/s	bf16	51.65 GB	~45.4	~181.6 q4
AMD Instinct MI250X	128.0 GB	3200.0 GB/s	bf16	51.65 GB	~43.4	~173.5 q4
Google TPU v5p	95.0 GB	2765.0 GB/s	bf16	51.65 GB	~37.5	~149.9 q4
NVIDIA A100 80GB SXM	80.0 GB	2039.0 GB/s	bf16	51.65 GB	~27.6	~110.5 q4
NVIDIA H100 PCIe	80.0 GB	2000.0 GB/s	bf16	51.65 GB	~27.1	~108.4 q4
NVIDIA GeForce RTX 5090	32.0 GB	1792.0 GB/s	q6	19.37 GB	~64.8	~97.2 q4
Google TPU v6e (Trillium)	32.0 GB	1638.0 GB/s	q6	19.37 GB	~59.2	~88.8 q4
AMD Instinct MI210	64.0 GB	1600.0 GB/s	fp8	25.82 GB	~43.4	~86.7 q4
NVIDIA A100 40GB SXM	40.0 GB	1555.0 GB/s	fp8	25.82 GB	~42.2	~84.3 q4
Google TPU v4	32.0 GB	1200.0 GB/s	q6	19.37 GB	~43.4	~65.1 q4
NVIDIA GeForce RTX 4090	24.0 GB	1008.0 GB/s	q5	16.14 GB	~43.7	~54.6 q4
AMD Radeon RX 7900 XTX	24.0 GB	960.0 GB/s	q5	16.14 GB	~41.6	~52.0 q4
NVIDIA GeForce RTX 3090	24.0 GB	936.0 GB/s	q5	16.14 GB	~40.6	~50.7 q4
NVIDIA L40S	48.0 GB	864.0 GB/s	fp8	25.82 GB	~23.4	~46.8 q4
Apple M3 Ultra	512.0 GB	819.0 GB/s	bf16	51.65 GB	~11.1	~44.4 q4
AMD Radeon RX 7900 XT	20.0 GB	800.0 GB/s	q4	12.91 GB	~43.4	~43.4 q4
Apple M2 Ultra	192.0 GB	800.0 GB/s	bf16	51.65 GB	~10.8	~43.4 q4
Apple M4 Max	128.0 GB	546.0 GB/s	bf16	51.65 GB	~7.4	~29.6 q4
Apple M1 Max	64.0 GB	400.0 GB/s	fp8	25.82 GB	~10.8	~21.7 q4
Apple M2 Max	96.0 GB	400.0 GB/s	bf16	51.65 GB	~5.4	~21.7 q4
Apple M3 Max	128.0 GB	400.0 GB/s	bf16	51.65 GB	~5.4	~21.7 q4
NVIDIA L4	24.0 GB	300.0 GB/s	q5	16.14 GB	~13.0	~16.3 q4
Apple M4 Pro	64.0 GB	273.0 GB/s	fp8	25.82 GB	~7.4	~14.8 q4
Apple M4	32.0 GB	120.0 GB/s	q6	19.37 GB	~4.3	~6.5 q4

Where it runs

Taken from the card's availability section.

Platform	Id on that platform
Hugging Face	google/diffusiongemma-26B-A4B-it

Verified benchmark evidence

Each row was checked against its source by a reviewer, and carries the model identifier as actually evaluated — which is not always the same as this card's.

Benchmark	Evaluated as	Score	Evidence date	Source kind
aa_lcr	DiffusionGemma 26B A4B	19.67%	2026-09-10 evaluated	independent evaluator	source
critpt	DiffusionGemma 26B A4B	0.29%	2026-09-10 evaluated	independent evaluator	source
gdpval_aa	DiffusionGemma 26B A4B	0.06%	2026-09-10 evaluated	independent evaluator	source
gpqa_diamond	DiffusionGemma 26B A4B	66.87%	2026-09-10 evaluated	independent evaluator	source

Reported benchmark scores

This card reports no benchmark scores yet.

Benchmark	Catalogue standing	Score

Data

This card as JSON · See it in the graph · Edit on GitHub