models · google

Gemma 4 26B-A4B

Google Gemma 4 MoE variant. 26B total / 4B active parameters. 262K context. Efficient inference with 115 tok/s throughput.

Compare Verify

What this registry has checked

Quarantine. We hold the publisher’s own published hash for these weights and have recorded it against this entry. We have not downloaded and hashed the weights ourselves, so the hash is the publisher’s claim, not our measurement.

Weight file SHA-256	25f5f6945c06b525d613f31a32c53d449de0e004f03d75ab5ddeb7a4d5823a48
Method	hf-published-sha256

Re-check this yourself → Source → Full record →

What is recorded

Publisher	google
Origin	United States
Licence	Apache-2.0
Parameters	25,805,936,206 (26B-A4B)
Context window	262K
Modality	text+image+video->text
Native precision	BF16
Pulls recorded here	180,015
Downloads reported upstream	48,151

Reported scores

GPQA Diamond	79.2
IFBench	72.4
Humanity's Last Exam	19.3

Reported by a third party and recorded here, not re-run by us. Source: artificialanalysis.ai/api/v2.

Measured on real hardware

None of these runs are ours. Each figure is the reporter’s, linked to their own evidence, under their own trust label. The number shown is single-stream decode — one user, concurrency 1 — because aggregate throughput across many users is a different quantity and reads as far faster than anyone will see.

Reported to local-ai-registry

Hardware	Engine	Build measured	Decode, 1 stream	Label	Reported by
GeForce RTX 5090	llama.cpp	Q3_K_M	256 tok/s @ 2,048	candidate	localmaxxing
Intel Arc Pro B70	llama.cpp	Q8_K_XL	125 tok/s @ 32,768	candidate	localmaxxing
Radeon PRO W7800	llama.cpp	Q4_K_S	117 tok/s @ 8,192	validated	0xsero
Intel Arc Pro B70	llama.cpp	UD-Q8_K_XL	115 tok/s @ 32,768	candidate	localmaxxing
Radeon PRO W7800	llama.cpp	UD-Q4_K_XL	113 tok/s @ 8,192	validated	0xsero
GeForce RTX 5060 Ti	llama.cpp	Q3_K_M	106 tok/s @ 2,048	candidate	localmaxxing
Apple M5 Max 128GB	mlx	MLX 4bit	102 tok/s @ 8,256	candidate	exo-postgres
GeForce RTX 5060 Ti	llama.cpp	Q3_K_XL	101 tok/s @ 2,048	candidate	localmaxxing
Radeon AI PRO R9700	llama.cpp	Q4_K_XL	98.0 tok/s @ 2,048	candidate	localmaxxing
Apple M4 Max 128GB	mlx	MLX 4bit	92.3 tok/s @ 8,256	candidate	exo-postgres

From local-ai-registry (MIT), commit 124959d: 58 runs across 23 machines, 8 validated — model revision and runtime pinned, launch accepted — and 50 candidate, their label for useful evidence without a reproducible-launch promise. 31 more not listed.

Measured by local.ai

Hardware	Build measured	Decode @ 8K	Memory @ 8K	Decode mode
NVIDIA GeForce RTX 5090 32GB	NVFP4	147 tok/s	29.4 GB	ordinary
NVIDIA RTX PRO 6000 Blackwell 96GB	NVFP4	139 tok/s	89.8 GB	ordinary
NVIDIA RTX 6000 Ada 48GB	NVFP4	116 tok/s	44.2 GB	ordinary
NVIDIA GeForce RTX 4090 24GB	NVFP4	114 tok/s	21.6 GB	ordinary
NVIDIA DGX Spark 128GB	NVFP4	27.7 tok/s	111.3 GB	ordinary
NVIDIA GeForce RTX 5090 32GB	NVFP4	247 tok/s	29.3 GB	not stated
NVIDIA GeForce RTX 5090 32GB	NVFP4	247 tok/s	29.3 GB	not stated
NVIDIA RTX PRO 6000 Blackwell 96GB	NVFP4	246 tok/s	90.8 GB	not stated

Measured by local.ai, not by us: 41 runs, 33 more not listed. Each names its engine pinned by image digest and the exact serve command, in the full record.

Elsewhere in the registry

All models → Browse with filters → What will run on my hardware? →

Registry record as of 2026-09-26. Identity, licence and parameter count are properties of a release and do not move; tier can, and the verify link above re-checks it against the live registry rather than this page.