Skip to content
Models / Gemma 4 26B A4B

Gemma 4 26B A4B

MoE enthusiast
HuggingFace → HuggingBay → Google → Cost & VRAM calculator → Find for my hardware

25.2B total, 3.8B active per token (MoE: 128 experts, top-8 + 1 shared). Multimodal (text + image). The small-MoE sweet spot - near-4B decode speed with 25B-class quality. NVFP4 (~14GB) needs NVIDIA Blackwell (vLLM) or Apple Silicon (MLX); Q4_K_M (~13GB weights, ~18GB runtime) for Ollama/llama.cpp.

AI-generated content marks

The provider reports that this model does not add embedded watermarks or provenance metadata to generated output.

Provider transparency docs →

coding reasoning chat general vision
Parameters
25.2B
Context
256k
License
other
Developer
Google
Origin
🇺🇸 USA
Released
Apr 2026

Scores

Coding
85
Reasoning
88
General
84

Score per dollar

5600 pts per $/M input

general_score (84) divided by cheapest input price ($0.02/M). Higher is better value. See live pricing.

Related models

Gemma 4 31B
30.7B enthusiast 19GB min RAM $0.10/M in Apr 2, 2026
Gemma 3 270M
270M edge 1GB min RAM Aug 14, 2025
Gemma 3n E2B
2.0B edge 6GB min RAM Jun 26, 2025
Gemma 3n E4B
4.0B edge 8GB min RAM Jun 26, 2025
Gemma 4 12B
12.0B consumer 7GB min RAM Jun 1, 2025
Gemma 3 12B
12.2B consumer 8GB min RAM $0.08/M in Feb 27, 2025
Gemma 3 1B
1.0B edge 1GB min RAM Feb 27, 2025
Gemma 3 27B
27.2B consumer 17GB min RAM $0.08/M in Feb 27, 2025
Gemma 3 4B
4.3B edge 3GB min RAM $0.04/M in Feb 27, 2025

Guides covering Gemma 4 26B A4B

How AT&T cut AI coding costs 56% with open-weight models and LiteLLM Aug 28, 2026 DGX Spark benchmarks: which local model wins in 2026 Aug 19, 2026

Save your hardware and every model page answers the real question: will it run on your machine, and how fast?

Join free - save your rig →

Run it locally

Per-quant memory needs and a static "can you run it?" reference - no rig entry required

NVFP4
14.0GB 15.0GB min 18.0GB rec
NVIDIA FP4 - Blackwell or Apple MLX, near-lossless
Q4_K_M
13.0GB 16.0GB min 18.0GB rec
Balanced - the usual local sweet spot
Q8_0
25.0GB 27.0GB min 30.0GB rec
Near-lossless
BF16
50.0GB 52.0GB min 56.0GB rec
Full quality, largest

The reference hardware

Schematic of the NVIDIA Jetson Orin NX 16GB reference rig - 16GB unified memory, 102 GB/s aggregate bandwidth
Schematic of the Single GTX 1080 Ti (11GB) reference rig - 11GB VRAM, 484 GB/s aggregate bandwidth
Schematic of the 4x H100 80GB (320GB) reference rig - 320GB VRAM, 13400 GB/s aggregate bandwidth
Schematic of the NVIDIA DGX Station 748GB reference rig - 748GB unified memory, 8000 GB/s aggregate bandwidth
Schematic of the 8x RTX 3090 rack (192GB) reference rig - 192GB VRAM, 7489 GB/s aggregate bandwidth
Schematic of the 4x RTX 5090 (128GB) reference rig - 128GB VRAM, 7168 GB/s aggregate bandwidth
Schematic of the AMD Instinct MI300X (192GB) reference rig - 192GB VRAM, 5324 GB/s aggregate bandwidth
Schematic of the 4x RTX 4090 (96GB) reference rig - 96GB VRAM, 4032 GB/s aggregate bandwidth
Schematic of the 2x RTX 5090 (64GB) reference rig - 64GB VRAM, 3584 GB/s aggregate bandwidth
Schematic of the 2x RTX 3090 (48GB) reference rig - 48GB VRAM, 1872 GB/s aggregate bandwidth
Schematic of the Single RTX 5090 (32GB) reference rig - 32GB VRAM, 1792 GB/s aggregate bandwidth
Schematic of the RTX PRO 6000 Blackwell (96GB) reference rig - 96GB VRAM, 1792 GB/s aggregate bandwidth
Schematic of the Mac Studio M4 Ultra 192GB reference rig - 192GB unified memory, 1092 GB/s aggregate bandwidth
Schematic of the Mac Studio M4 Ultra 512GB reference rig - 512GB unified memory, 1092 GB/s aggregate bandwidth
Schematic of the Single RTX 4090 (24GB) reference rig - 24GB VRAM, 1008 GB/s aggregate bandwidth
Schematic of the MacBook Pro M5 Max 128GB reference rig - 128GB unified memory, 614 GB/s aggregate bandwidth
Schematic of the Dual EPYC 9004 + 768GB DDR5-4800 reference rig - 768GB unified memory, 460 GB/s aggregate bandwidth
Schematic of the DGX Spark 128GB unified reference rig - 128GB unified memory, 273 GB/s aggregate bandwidth
Schematic of the Ryzen AI Max+ 395 128GB reference rig - 128GB unified memory, 256 GB/s aggregate bandwidth
Schematic of the Jetson AGX Orin 64GB reference rig - 64GB unified memory, 204 GB/s aggregate bandwidth
Schematic of the Epyc + 512GB DDR4-3200 + 2x RTX 3090 reference rig - 560GB unified memory, 204 GB/s aggregate bandwidth
Schematic of the Epyc + 512GB DDR4-2400 + 2x RTX 3090 reference rig - 560GB unified memory, 153 GB/s aggregate bandwidth

22 reference configs, drawn in-house. Scroll for more.

Can you run it? - reference rigs

Rig NVFP4 Q4_K_M Q8_0 BF16
NVIDIA Jetson Orin NX 16GB no -> cloud tight no -> cloud no -> cloud
Single GTX 1080 Ti (11GB) no -> cloud offload no -> cloud no -> cloud
4x H100 80GB (320GB) no -> cloud fast 1526.4t/s fast 1110.3t/s fast 708.1t/s
NVIDIA DGX Station 748GB no -> cloud fast 911.3t/s fast 662.9t/s fast 422.8t/s
8x RTX 3090 rack (192GB) no -> cloud fast 853.2t/s fast 620.6t/s fast 395.8t/s
4x RTX 5090 (128GB) fast 791.8t/s fast 816.5t/s fast 593.9t/s fast 378.8t/s
AMD Instinct MI300X (192GB) no -> cloud fast 606.6t/s fast 441.2t/s fast 281.4t/s
4x RTX 4090 (96GB) no -> cloud fast 459.3t/s fast 334.1t/s fast 213.1t/s
2x RTX 5090 (64GB) fast 395.9t/s fast 408.3t/s fast 297.0t/s fast 189.4t/s
2x RTX 3090 (48GB) no -> cloud fast 213.3t/s fast 155.1t/s offload
Single RTX 5090 (32GB) fast 198.0t/s fast 204.1t/s fast 148.5t/s offload
RTX PRO 6000 Blackwell (96GB) fast 198.0t/s fast 204.1t/s fast 148.5t/s fast 94.7t/s
Mac Studio M4 Ultra 192GB fast 131.6t/s fast 135.7t/s fast 98.7t/s fast 63.0t/s
Mac Studio M4 Ultra 512GB fast 131.6t/s fast 135.7t/s fast 98.7t/s fast 63.0t/s
Single RTX 4090 (24GB) no -> cloud fast 114.8t/s offload no -> cloud
MacBook Pro M5 Max 128GB fast 74.0t/s fast 76.3t/s fast 55.5t/s fast 35.4t/s
Dual EPYC 9004 + 768GB DDR5-4800 no -> cloud fast 52.5t/s fast 38.2t/s fast 24.4t/s
DGX Spark 128GB unified fast 30.2t/s fast 31.1t/s fast 22.6t/s ok 14.4t/s
Ryzen AI Max+ 395 128GB no -> cloud fast 29.2t/s fast 21.2t/s ok 13.5t/s
Jetson AGX Orin 64GB no -> cloud fast 23.3t/s ok 17.0t/s ok 10.8t/s
Epyc + 512GB DDR4-3200 + 2x RTX 3090 no -> cloud fast 23.3t/s ok 17.0t/s ok 10.8t/s
Epyc + 512GB DDR4-2400 + 2x RTX 3090 no -> cloud ok 17.5t/s ok 12.7t/s ok 8.1t/s

Fit tiers use the same will-it-run logic as the rig finder. For comfortable fits, the badge reflects decode speed: fast >=20 t/s, ok 8-20 t/s, slow <8 t/s. t/s is a bandwidth estimate, not a measured benchmark.

Download options

NVFP4 official -1% vs fp16
14.0GB dl 15.0GB min 18.0GB rec
REC RAM vs largest quant
NVFP4 - NVIDIA Blackwell via vLLM (nvidia/Gemma-4-26B-A4B-NVFP4) or Apple Silicon via MLX (ollama run gemma4:26b-nvfp4). Not supported on Ada/Ampere/Hopper/AMD.
GGUF on HF →
Q4_K_M official -5% vs fp16
13.0GB dl 16.0GB min 18.0GB rec
REC RAM vs largest quant
Q8_0 official -1% vs fp16
25.0GB dl 27.0GB min 30.0GB rec
REC RAM vs largest quant
BF16 official
50.0GB dl 52.0GB min 56.0GB rec
REC RAM vs largest quant
GGUF on HF →

Or run it in the cloud

Live per-provider pricing, throughput and uptime - refreshed about 10 hours ago via OpenRouter. Click a column to sort.

Provider Type Input $/M Output $/M Cache $/M Tok/s Latency Uptime Value
PrimeIntellect
API 0.02 0.04 0.007 - - 100.00% best uptime
Darkbloom
API 0.04 0.15 0.010 - - 100.00%
DeepInfra
API 0.07 0.34 0.042 - - 100.00%
AkashML
API 0.07 0.24 0.046 - - 100.00%
NextBit
API 0.08 0.26 0.042 - - 100.00%
Io Net
API 0.08 0.28 0.037 - - 100.00%
Makora
API 0.09 0.28 0.049 - - 100.00%
Cloudflare
API 0.10 0.30 0.050 - - 100.00%
CoreWeave
API 0.10 0.30 0.050 - - 100.00%
Novita
API 0.13 0.40 0.042 - - 100.00%
Parasail
API 0.13 0.40 0.050 - - 100.00%
Venice
API 0.13 0.40 0.050 - - 100.00%
SiliconFlow
API 0.14 0.40 0.050 - - 100.00%
Google
API 0.15 0.60 0.042 - - 100.00%
DekaLLM avoid
API 0.06 0.33 0.040 - - 46.67%

Default order: throughput among 95%+ uptime providers, then latency; subscriptions last. Sort by any column. Subscription rows show $/mo in the Value column - per-token columns are "-". Affiliate links are marked sponsored / nofollow. Confirm current pricing on the provider's site before committing.

Detailed API pricing page + JSON endpoint →

See who runs Google in production →

PRICE HISTORY

Inference cost over time

Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.

Loading price history...

Tokenstead community

Need a second opinion on this model?

Ask the people running it on real hardware, then post your tok/s and fit notes so the next reader has a data point.

Join the community Share your rig

Something wrong or stale? Send a correction.