ℹ

Public note: Manuell erfasste Messung (fairer Side-by-Side ai-001 vs. ai-002, 2026-09-26). Single-Stream, 256 Tokens, greedy -> ~70 tok/s Generation. Hardware: 1x RTX PRO 6000 Blackwell (Full Workstation, 600W-Profil, real ~391W), TP=1, CUDA-Graphs an (kein enforce-eager), gpu-memory-utilization 0.90, ctx 131072, max-num-seqs 10. WICHTIG: Der System-RAM lief mit 3000 MHz (2 Riegel). Variante: Uncensored (mazinb).

Generation70,00tok/s

Prefill0,00tok/s

Time to First Token0,00ms

Total duration3,66s

Concurrency1parallel

Ranking in the field

This GPU All GPUs

55of 86 systems

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

This run is better than 36 % of all comparable systems.

Generation 70,0 tok/s

-51 % vs Ø 144,3

Prefill 0,0 tok/s

-100 % vs Ø 3.968,6

Time to First Token 0 ms

-100 % vs Ø 15.931

Distribution in the field0 – 395 tok/s

Ø 144 Median Dieser Lauf

Wie schlägt sich dieser Benchmark mit anderen Modellen?

gpt-oss-20bNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260730-035052-9f766a

395,2 tok/s

Nemotron-3-Nano-4BNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260727-181607-bffb11

393,5 tok/s

gemma-4-E2B-itNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260727-181606-6ea089

378,6 tok/s

Nemotron-3-Nano-Omni-30B-A3B-ReasoningNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260729-032119-4852bf

357,1 tok/s

Nemotron-3-Nano-30B-A3BNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260729-032121-643b00

357,0 tok/s

Nemotron-Cascade-2-30B-A3BNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260729-032120-4e3476

356,8 tok/s

Qwen3-30B-A3B-Thinking-2507NVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260729-032123-b33938

312,1 tok/s

Qwen3-Coder-30B-A3B-InstructNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260729-032124-ac5167

311,3 tok/s

Laguna-XS-2.1NVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260728-140954-cecd39

310,2 tok/s

Laguna-XS-2.1NVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260728-140955-42efc5

309,9 tok/s

Laguna-XS-2.1NVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260728-140955-abb113

309,5 tok/s

Qwen3-30B-A3B-Instruct-2507NVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260729-032121-b5f531

304,6 tok/s

Qwen3-Omni-30B-A3B-ThinkingNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260729-032125-29dbef

304,4 tok/s

Qwen3-VL-30B-A3B-InstructNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260729-032125-021dfb

303,6 tok/s

Qwen3.8-Flash-Next this runNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20261023-184710-4b5edd

70,0 tok/s

How does this benchmark compare on other GPUs?

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

NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20261012-182514-5f16dc

164,6 tok/s

NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20261012-175110-537c3d

158,1 tok/s

NVIDIA RTX A6000run-20261021-093815-fdc402

125,2 tok/s

NVIDIA GeForce RTX 3090run-20261013-092430-8e3577

116,4 tok/s

NVIDIA RTX A6000run-20261021-093816-efa091

115,5 tok/s

NVIDIA RTX PRO 6000 Blackwell Workstation Editionrun-20261023-180245-bd93b8

96,7 tok/s

NVIDIA RTX PRO 6000 Blackwell Workstation Editionrun-20261023-184710-e3cffa

95,0 tok/s

NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20261023-183307-2a60cf

79,1 tok/s

2× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20261023-180239-49eb27

77,2 tok/s

2× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20261023-180238-ae71f7

74,2 tok/s

NVIDIA RTX PRO 6000 Blackwell Workstation Edition this runrun-20261023-184710-4b5edd

70,0 tok/s

NVIDIA Tesla V100 PCIe 32GBrun-20261021-093816-154b4e

59,3 tok/s

NVIDIA Tesla V100 PCIe 32GBrun-20261019-134138-464b84

52,1 tok/s

NVIDIA GeForce RTX 3090run-20261013-165111-42c54a

50,5 tok/s

Hardware

GPU: NVIDIA RTX PRO 6000 Blackwell Workstation Edition · 96 GB VRAM
CPU: AMD Ryzen 9 9950X 16-Core Processor
RAM: 123 GB
Mainboard: ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI

Setup

Runtime: vLLM
Quantization: NVFP4
Model: Qwen3.8-Flash-Next

Configuration

benchmark-konfiguration — run-20261023-184710-4b5edd

# LLM-Benchmark Konfiguration # Modell : Qwen3.8-Flash-Next # Engine : vLLM # Run-ID : run-20261023-184710-4b5edd # GPU : NVIDIA RTX PRO 6000 Blackwell Workstation Edition # CPU : AMD Ryzen 9 9950X 16-Core Processor # RAM : 123 GB bench@llm-benchmark:~$ vllm serve \ --model mazinb/Qwen3.8-Flash-Next-Uncensored-NVFP4 \ --trust-remote-code \ --distributed-executor-backend mp \ --gpu-memory-utilization 0.90 \ --max-model-len 131072 \ --max-num-seqs 10

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
Modellalias?Der Name, unter dem das Modell ueber die API angesprochen wird. Genau dieser Wert muss im Request-Feld 'model' stehen.	Qwen3.8-Flash-Next-Uncensored
Kontextlaenge?Maximale Anzahl Tokens (Eingabe + erzeugte Ausgabe zusammen), die das Modell pro Anfrage verarbeiten kann.	131072
Executor-Backend?Verfahren fuer Multi-GPU/Multi-Node: mp = Prozesse auf einer Maschine, ray = ueber Ray (auch ueber mehrere Knoten).	mp
GPU-Speicher?Anteil des GPU-Speichers (0 bis 1), den vLLM belegen darf. 0.92 = 92 %. Hoeher = mehr Platz fuer den KV-Cache (mehr/laengere parallele Anfragen), aber groesseres Risiko fuer 'Out of Memory'.	0.90
Kontext?Groesse des Kontextfensters in Token. 0 = der beim Training verwendete Kontext des Modells.	131072
Max-Sequenzen?Maximale Anzahl gleichzeitig bearbeiteter Anfragen (Batch-Groesse). Hoeher = mehr Durchsatz, braucht aber mehr KV-Cache-Speicher.	10
execution_type	local

All benchmarks of this model To leaderboard

Model comparison

Qwen3.8-Flash-Next 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

NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 1.564,6 tok/sNVIDIA RTX A6000 1.171,8 tok/sNVIDIA GeForce RTX 3090 1.132,7 tok/s★ NVIDIA RTX PRO 6000 Blackwell Workstation Edition 584,1 tok/s this runNVIDIA GB10 (DGX Spark) 492,8 tok/sNVIDIA Tesla V100 PCIe 32GB 196,2 tok/sAMD Radeon RX 7900 XTX 18,7 tok/s

CPUby processor

AMD Ryzen 9 7945HX with Radeon Graphics 1.564,6 tok/sAMD Ryzen Threadripper PRO 3945WX 12-Cores 1.171,8 tok/sAMD Ryzen 5 5600X 6-Core Processor 1.132,7 tok/s★ AMD Ryzen 9 9950X 16-Core Processor 584,1 tok/s this runNVIDIA Grace 492,8 tok/sAMD Ryzen Threadripper PRO 9965WX 24-Cores 424,2 tok/sAMD Ryzen 3 PRO 3200GE w/ Radeon Vega Graphics 196,2 tok/sAMD EPYC 7203P 8-Core Processor 79,2 tok/sIntel(R) Xeon(R) CPU E5-4657L v2 @ 2.40GHz 63,2 tok/sAMD Ryzen Threadripper PRO 3955WX 16-Cores 18,7 tok/s

MBby mainboard

Shenzhen Meigao Electronic Equipment Co.,Ltd DRFXI (MotherBoard Series) 1.564,6 tok/sunbekannt 1.171,8 tok/sASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING 1.132,7 tok/s★ ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI 584,1 tok/s this runASUSTeK COMPUTER INC. GX10 492,8 tok/sASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE 424,2 tok/sASUSTeK COMPUTER INC. ROG STRIX X570-F GAMING 196,2 tok/sASRockRack ROMED8-2T/BCM 79,2 tok/sDell Inc. PowerEdge R820 63,2 tok/sASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 35,4 tok/s

ENGby engine

strata 1.564,6 tok/sllama.cpp 1.555,3 tok/sunbekannt 1.171,8 tok/s★ vLLM 584,1 tok/s this run

DRVby driver

unbekannt 1.564,6 tok/sNVIDIA 590.48.01 / CUDA 13.1 492,8 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 URL Reset

⚡ Electricity price EUR/kWh

⚙️ System utilization 100 %

🖥️ Acquisition EUR

🔌 Idle 70 W

⚡ TDP 644 W

☁️ External LLM (API) show

Input €/1M Output €/1M Input per output token

Electricity0.30 EUR/kWh

Avg power (incl. idle)644 W estimated (TDP)GPU 600 + CPU 29 + Board 15 W full load

Avg cost / hourEUR 0.19

Electricity / 1M tokensEUR 0.77

Token / kWh391.46K

Acquisition (system)EUR 15,558 full priceGPU EUR 13,000 · CPU EUR 649 · Board EUR 499 · RAM EUR 1,230 · PSU EUR 180

Electricity (2 years)–

TCO (2 years)EUR 18,942

Output tokens (2 years)4.42B

☁️ 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 (70 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

	Qwen3.8-Flash-NextNVIDIA RTX PRO 6000 Blackwell Workstation Edition	Qwen3.8-Flash-NextNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition	Qwen3.8-Flash-NextNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition	Qwen3.8-Flash-NextNVIDIA RTX A6000
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)	–	–	–	–

NVIDIA RTX PRO 6000 Blackwell Workstation Edition · this run 70 tok/s NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 165 tok/s NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 158 tok/s NVIDIA RTX A6000 125 tok/s

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.