Gemma 4 26B-A4B on NVIDIA Grace Blackwell Ultra GB300

Yes — Grace Blackwell Ultra GB300 runs Gemma 4 26B-A4B excellently at Q8_0 — 500 tok/s. 288 GB VRAM with plenty of headroom.

Model Size

25.2B

Device VRAM

288 GB

Bandwidth

8000 GB/s

Quantization

Q8_0

Benchmarks

Performance by Quantization

OwnRig currently has one published compatibility entry for Gemma 4 26B-A4B on NVIDIA Grace Blackwell Ultra GB300 at Q8_0. This is the best supported pairing we can stand behind today.

Quantization	Speed	TTFT	Fits in VRAM	Rating	Confidence
Q8_0	500 tok/s	40ms	✓ Yes	Excellent	estimated

Notes

Q8_0

25.2B model. Q8_0 26.86GB on 288GB discrete_gpu.

About Gemma 4 26B-A4B

Gemma 4 26B-A4B (25.2B) is a chat, coding, reasoning, multi-purpose model. Mixture-of-Experts architecture: 25.2B total parameters but only 3.8B active per token (8 selected + 1 shared expert per layer, out of 128 total). Hybrid dense+sparse FFN design. Inference throughput closer to a 4B dense model; quality closer to a 27B dense model. 256K context window. Benchmarks: 88.3% AIME 2026, 82.6% MMLU Pro, 77.1% LiveCodeBench. All 25.2B weights must be loaded into VRAM despite sparse activation; fits on 24 GB GPUs at Q4_K_M. Apache 2.0 licensed.

View all Gemma 4 26B-A4B hardware options →

About NVIDIA Grace Blackwell Ultra GB300

NVIDIA Grace Blackwell Ultra GB300 has 288 GB at 8000 GB/s. Street price: $30,000.

See all models NVIDIA Grace Blackwell Ultra GB300 can run →

Estimate method: Estimated from model architecture, quantization size, and device bandwidth. Reference hardware source: huggingface.co (2026-04-04)

Performance varies by driver version, inference engine, quantization method, context length, and system configuration. Figures shown are estimates based on community benchmarks and may not reflect your exact setup. Product names are trademarks of their respective owners. OwnRig is independent and not affiliated with any hardware or AI model provider.