Created bymario-alka.dePowered bygodcore.denoob2claw.detricoma.de
Contributed byMario AlkaGoogle

Gemma-4-26B-A4B-it

Performance benchmark · measured on 27.08.2026 10:34

Benchmark-IDrun-20260827-090901-dbf3d0
Timebench 3 - Kombi (Prefill + Generation)MoE26BRuntime: vLLMQuantisierung: AWQ
Generation145,21tok/s
Prefill8.912,90tok/s
Time to First Token221,00ms
Total duration14,55s
Concurrency1parallel
Ranking in the field
215of 1300 systems

Performance benchmark · Primary metric: Generation-Speed (tok/s) · 1× concurrent

This run is better than 84 % of all comparable systems.
Generation 145,2 tok/s
+91 % vs Ø 76,2
Prefill 8.912,9 tok/s
+245 % vs Ø 2.582,6
Time to First Token 221 ms
-99 % vs Ø 31.940
Distribution in the field0 – 405 tok/s
Ø 76 Median Dieser Lauf

Wie schlägt sich dieser Benchmark mit anderen Modellen?

How does this benchmark compare on other GPUs?

Same model on different hardware · 1× concurrent · Generation (tok/s)

Hardware

GPU: 2x NVIDIA RTX A6000 · 48 GB VRAM
CPU: AMD Ryzen Threadripper PRO 7955WX 16-Cores
RAM: 184 GB
Mainboard: ASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE

Setup

Runtime: vLLM
Quantization: AWQ
Model: Gemma-4-26B-A4B-it

Configuration

benchmark-konfiguration — run-20260827-090901-dbf3d0
# LLM-Benchmark Konfiguration # Modell : Gemma-4-26B-A4B-it # Engine : vLLM # Run-ID : run-20260827-090901-dbf3d0 # GPU : 2x NVIDIA RTX A6000 # CPU : AMD Ryzen Threadripper PRO 7955WX 16-Cores # RAM : 184 GB bench@llm-benchmark:~$ vllm \ --tensor-parallel-size 2 \ --max-model-len 16384 '(2x' RTX A6000 '48GB)'
Engine?Die Inferenz-Software, die das Modell ausliefert (z.B. vLLM oder llama.cpp). Sie bestimmt Geschwindigkeit, unterstuetzte Modellformate und welche Parameter ueberhaupt verfuegbar sind.vllm
Kontextlaenge?Maximale Anzahl Tokens (Eingabe + erzeugte Ausgabe zusammen), die das Modell pro Anfrage verarbeiten kann.16384
Tensor-Parallel?Anzahl GPUs, auf die JEDER einzelne Modell-Layer aufgeteilt wird (Tensor-Parallelitaet). Mehr GPUs = mehr VRAM und meist mehr Speed, aber die Anzahl der Attention-Heads muss durch diesen Wert teilbar sein (z.B. 32 Heads -> nur 1, 2, 4, 8 ... moeglich, NICHT 3).2
Kontext?Groesse des Kontextfensters in Token. 0 = der beim Training verwendete Kontext des Modells.16384

All benchmarks of this model To leaderboard

Anzeige
Model comparison

Gemma-4-26B-A4B-it on various hardware

All published performance runs of this model – each bubble a variant: position = prefill (X) × generation (Y), bubble size = number of runs. Closer to the top right = faster. ★ Marked gold = this benchmark.

GPUby graphics card

2.0501.5371.0255120,004.2438.48612.730Prefill (tok/s)Generation (tok/s)NVIDIA RTX PRO 6000 Blackwell Workstation Edition - 1.589,1 tok/s Generation, 6.148 tok/s Prefill, TTFT 3.731 ms (3 Laufe)NVIDIA RTX PRO 6000 B...NVIDIA GeForce RTX 5090 - 1.540,5 tok/s Generation, 4.006 tok/s Prefill, TTFT 4.749 ms (3 Laufe)NVIDIA GeForce RTX 50...NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition - 1.479,9 tok/s Generation, 7.172 tok/s Prefill, TTFT 3.769 ms (9 Laufe)NVIDIA RTX PRO 6000 B...NVIDIA GeForce RTX 5070 Ti - 1.009,0 tok/s Generation, 3.578 tok/s Prefill, TTFT 5.953 ms (3 Laufe)NVIDIA GeForce RTX 50...NVIDIA GeForce RTX 3090 Ti - 926,2 tok/s Generation, 3.438 tok/s Prefill, TTFT 4.863 ms (6 Laufe)NVIDIA GeForce RTX 30...AMD Radeon 8060S Graphics - 79,5 tok/s Generation, 533 tok/s Prefill, TTFT 9.976 ms (9 Laufe)AMD Radeon 8060S Grap...NVIDIA RTX A6000 - 754,7 tok/s Generation, 10.343 tok/s Prefill, TTFT 1.837 ms (18 Laufe) | DIESER LAUF★ NVIDIA RTX A6000
NVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.589,1 tok/sNVIDIA GeForce RTX 5090 1.540,5 tok/sNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 1.479,9 tok/sNVIDIA GeForce RTX 5070 Ti 1.009,0 tok/sNVIDIA GeForce RTX 3090 Ti 926,2 tok/s★ NVIDIA RTX A6000 754,7 tok/s this runAMD Radeon 8060S Graphics 79,5 tok/s

CPUby processor

2.0501.5371.0255120,004.2438.48612.730Prefill (tok/s)Generation (tok/s)AMD Ryzen 9 9950X 16-Core Processor - 1.589,1 tok/s Generation, 6.148 tok/s Prefill, TTFT 3.731 ms (3 Laufe)AMD Ryzen 9 9950X 16-...AMD Ryzen 7 5800X3D 8-Core Processor - 1.540,5 tok/s Generation, 4.006 tok/s Prefill, TTFT 4.749 ms (3 Laufe)AMD Ryzen 7 5800X3D 8...AMD Ryzen Threadripper PRO 9965WX 24-Cores - 1.479,9 tok/s Generation, 7.172 tok/s Prefill, TTFT 3.769 ms (9 Laufe)AMD Ryzen Threadrippe...AMD Ryzen Threadripper PRO 5975WX 32-Cores - 1.009,0 tok/s Generation, 3.578 tok/s Prefill, TTFT 5.953 ms (3 Laufe)AMD Ryzen Threadrippe...AMD Ryzen 9 8945HX with Radeon Graphics - 926,2 tok/s Generation, 3.620 tok/s Prefill, TTFT 6.478 ms (3 Laufe)AMD Ryzen 9 8945HX wi...AMD Ryzen 5 5600X 6-Core Processor - 397,2 tok/s Generation, 3.255 tok/s Prefill, TTFT 3.248 ms (3 Laufe)AMD Ryzen 5 5600X 6-C...AMD RYZEN AI MAX+ 395 w/ Radeon 8060S - 79,5 tok/s Generation, 533 tok/s Prefill, TTFT 9.976 ms (9 Laufe)AMD RYZEN AI MAX+ 395...AMD Ryzen Threadripper PRO 7955WX 16-Cores - 754,7 tok/s Generation, 10.343 tok/s Prefill, TTFT 1.837 ms (18 Laufe) | DIESER LAUF★ AMD Ryzen Threadrippe...
AMD Ryzen 9 9950X 16-Core Processor 1.589,1 tok/sAMD Ryzen 7 5800X3D 8-Core Processor 1.540,5 tok/sAMD Ryzen Threadripper PRO 9965WX 24-Cores 1.479,9 tok/sAMD Ryzen Threadripper PRO 5975WX 32-Cores 1.009,0 tok/sAMD Ryzen 9 8945HX with Radeon Graphics 926,2 tok/s★ AMD Ryzen Threadripper PRO 7955WX 16-Cores 754,7 tok/s this runAMD Ryzen 5 5600X 6-Core Processor 397,2 tok/sAMD RYZEN AI MAX+ 395 w/ Radeon 8060S 79,5 tok/s

MBby mainboard

2.0571.5431.0295140,003.8347.66811.502Prefill (tok/s)Generation (tok/s)ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI - 1.589,1 tok/s Generation, 6.148 tok/s Prefill, TTFT 3.731 ms (3 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING - 1.540,5 tok/s Generation, 4.006 tok/s Prefill, TTFT 4.749 ms (3 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI - 1.009,0 tok/s Generation, 3.578 tok/s Prefill, TTFT 5.953 ms (3 Laufe)ASUSTeK COMPUTER INC....Meigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) - 926,2 tok/s Generation, 3.620 tok/s Prefill, TTFT 6.478 ms (3 Laufe)Meigao Innovation Tec...ASUSTeK COMPUTER INC. PRIME A520M-K - 397,2 tok/s Generation, 3.255 tok/s Prefill, TTFT 3.248 ms (3 Laufe)ASUSTeK COMPUTER INC....Bosgame AXB35-02 (BeyondMax Series) - 79,5 tok/s Generation, 591 tok/s Prefill, TTFT 11.147 ms (8 Laufe)Bosgame AXB35-02 (Bey...Meigao Innovation Technology (Shen Zhen) Co., Ltd SHWSA (MS-S1 MAX) - 42,2 tok/s Generation, 71 tok/s Prefill, TTFT 609 ms (1 Lauf)Meigao Innovation Tec...ASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE - 1.479,9 tok/s Generation, 9.286 tok/s Prefill, TTFT 2.481 ms (27 Laufe) | DIESER LAUF★ ASUSTeK COMPUTER INC....
ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI 1.589,1 tok/sASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING 1.540,5 tok/s★ ASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE 1.479,9 tok/s this runASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 1.009,0 tok/sMeigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) 926,2 tok/sASUSTeK COMPUTER INC. PRIME A520M-K 397,2 tok/sBosgame AXB35-02 (BeyondMax Series) 79,5 tok/sMeigao Innovation Technology (Shen Zhen) Co., Ltd SHWSA (MS-S1 MAX) 42,2 tok/s

ENGby engine

1.9861.4909934970,006.46312.92619.389Prefill (tok/s)Generation (tok/s)llama.cpp - 1.589,1 tok/s Generation, 4.304 tok/s Prefill, TTFT 5.575 ms (38 Laufe)llama.cppunbekannt - 397,2 tok/s Generation, 2.459 tok/s Prefill, TTFT 2.588 ms (4 Laufe)unbekanntvLLM - 754,7 tok/s Generation, 15.993 tok/s Prefill, TTFT 780 ms (9 Laufe) | DIESER LAUF★ vLLM
llama.cpp 1.589,1 tok/s★ vLLM 754,7 tok/s this rununbekannt 397,2 tok/s

DRVby driver

2.0571.5431.0295140,002.6195.2377.856Prefill (tok/s)Generation (tok/s)unbekannt - 1.589,1 tok/s Generation, 6.345 tok/s Prefill, TTFT 4.573 ms (50 Laufe)unbekanntROCm 7.2.0 - 42,2 tok/s Generation, 71 tok/s Prefill, TTFT 609 ms (1 Lauf)ROCm 7.2.0
unbekannt 1.589,1 tok/sROCm 7.2.0 42,2 tok/s
💰 Economics

Economics of this run

Operating cost, TCO and comparison with the next-best runs of the same model at identical concurrency (1× concurrent). Methodology →

⚙️ ConfigurationAll metrics and charts below follow these settings – based on a 24-month runtime.Save to URLReset
⚡ Electricity price EUR/kWh
⚙️ System utilization 100 %
🖥️ Acquisition EUR
🔌 Idle 95 W
⚡ TDP 671 W
☁️ External LLM (API)
Electricity0.30 EUR/kWh
Avg power (incl. idle)671 W estimated (TDP)GPU 600 + CPU 46 + Board 25 W full load
Avg cost / hourEUR 0.20
Electricity / 1M tokensEUR 0.39
Token / kWh779.07K
Acquisition (system)EUR 11,454 full priceGPU EUR 6,000 · CPU EUR 1,399 · Board EUR 1,299 · RAM EUR 2,576 · PSU EUR 180
Electricity (2 years)
TCO (2 years)EUR 14,981
Output tokens (2 years)9.16B
☁️ External LLM (API) – comparison
External LLM cost (2 years)
Savings vs. external (2 years)

All values above and the charts below take the configured system utilization into account: at X% the system generates only X% of the time, the rest it idles (95 W). Cost per hour drops (more idle), cost per token rises.

Cost over 2 years – electricity only

Cost over 2 years – incl. acquisition (TCO)

Speed vs. tokens per euro

Euro per 1M tokens

Comparison vs. API – economics per benchmark

Gemma-4-26B-A4B-it2x NVIDIA RTX A6000Gemma-4-26B-A4B-itNVIDIA GeForce RTX 5090Gemma-4-26B-A4B-itNVIDIA RTX PRO 6000 Blackwell Workstation EditionGemma-4-26B-A4B-it3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition
Electricity cost (24 mo.)
Acquisition cost
Total cost (TCO)
Generated tokens (24 mo.)
Token price via API
Break-even point (days)
Result (savings / extra cost)

Comparison with up to 3 next-best runs of this model at the same concurrency (at least one on different hardware). Power = GPU TDP + CPU (idle + 15 %) + board (estimated), acquisition = full system (GPU + CPU + board + RAM + PSU), prices = stored market prices.

Contributed by

Mario Alka Administrator

@marioalka

Ich bin Unternehmer, Softwareentwickler und KI-Enthusiast. Seit vielen Jahren entwickle ich Unternehmenssoftware und beschäftige mich inzwischen fast täglich mit lokalen LLMs, KI-Agenten und leistungsfähiger KI-Hardware.

Mit LLM-Benchmark.de möchte ich eine Plattform schaffen, auf der Modelle, GPUs und Agenten objektiv und reproduzierbar miteinander verglichen werden.