ℹ

Public note: TensorFold 1.1.6 (CUDA-Backend) mit MLX-q4g64 (4 Bit, group_size 64) auf einer RTX 3090 (24 GB), Ryzen 5 5600X, 32 GB RAM. Das Modell liegt mit 19,09 GiB vollstaendig im VRAM, ohne CPU-Offload; Kontext 8192, 7 gleichzeitige Streams. Thinking per Chat-Template aktiv. Teil eines Engine-Vergleichs (vLLM / TensorFold) auf identischer Hardware und identischem Modell.

Generation145,72tok/s

Prefill6.272,56tok/s

Time to First Token307,50ms

Total duration14,67s

Concurrency1parallel

Ranking in the field

This GPU All GPUs

10of 18 systems

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

This run is better than 47 % of all comparable systems.

Generation 145,7 tok/s

-4 % vs Ø 151,4

Prefill 6.272,6 tok/s

+84 % vs Ø 3.412,8

Time to First Token 308 ms

-90 % vs Ø 2.945

Distribution in the field18 – 471 tok/s

Ø 151 Median Dieser Lauf

Wie schlägt sich dieser Benchmark mit anderen Modellen?

NVIDIA-Nemotron-3.5-Lightning-30B-A3BNVIDIA GeForce RTX 3090 · run-20261026-183943-a8c4f8

471,3 tok/s

NVIDIA-Nemotron-3.5-Lightning-30B-A3BNVIDIA GeForce RTX 3090 · run-20261026-180644-ba787b

235,7 tok/s

NVIDIA-Nemotron-3.5-Lightning-30B-A3BNVIDIA GeForce RTX 3090 · run-20261026-183943-18c236

218,8 tok/s

NVIDIA-Nemotron-3.5-Lightning-30B-A3BNVIDIA GeForce RTX 3090 · run-20261026-183943-804e04

211,4 tok/s

gpt-oss-20bNVIDIA GeForce RTX 3090 · run-20261024-084507-2c110a

211,1 tok/s

NVIDIA-Nemotron-3.5-Lightning-30B-A3BNVIDIA GeForce RTX 3090 · run-20261026-183943-3923eb

210,7 tok/s

Qwen3-30B-A3B-Thinking-2507NVIDIA GeForce RTX 3090 · run-20261024-084507-3fb87c

189,2 tok/s

NVIDIA-Nemotron-3.5-Lightning-30B-A3BNVIDIA GeForce RTX 3090 · run-20261026-183943-627b3c

176,0 tok/s

Qwen3.6-35B-A3B same modelNVIDIA GeForce RTX 3090 · run-20261024-084507-4aa5d8

172,4 tok/s

Qwen3.6-35B-A3B this runNVIDIA GeForce RTX 3090 · run-20261101-195359-a93dda

145,7 tok/s

Qwen3.6-35B-A3B same modelNVIDIA GeForce RTX 3090 · run-20261101-202817-9736b2

132,7 tok/s

Qwen3.8-Flash-NextNVIDIA GeForce RTX 3090 · run-20261013-092430-8e3577

116,4 tok/s

NVIDIA-Nemotron-3.5-Lightning-30B-A3BNVIDIA GeForce RTX 3090 · run-20261024-090453-79907e

101,0 tok/s

Qwen3.8-Flash-NextNVIDIA GeForce RTX 3090 · run-20261013-165111-42c54a

50,5 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 Workstation Editionrun-20260729-032129-9a4f52

267,1 tok/s

3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260731-071556-663768

243,4 tok/s

3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260731-071555-e55e07

228,7 tok/s

3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260731-071555-c7eec3

223,1 tok/s

NVIDIA GeForce RTX 5090run-20260731-071557-78f69d

208,5 tok/s

NVIDIA RTX PRO 6000 Blackwell Workstation Editionrun-20260724-083000-d5c2ec

195,0 tok/s

2× NVIDIA RTX A6000run-20260829-234039-49ac6d

172,5 tok/s

NVIDIA GeForce RTX 3090run-20261024-084507-4aa5d8

172,4 tok/s

NVIDIA GeForce RTX 3090 Tirun-20260729-105337-6a9108

164,0 tok/s

NVIDIA GeForce RTX 3090 Tirun-20260731-084840-1166d8

162,4 tok/s

NVIDIA GeForce RTX 3090 Tirun-20260804-052131-5d515f

162,2 tok/s

2× NVIDIA RTX A6000run-20260829-152523-570de6

156,0 tok/s

2× NVIDIA RTX A6000run-20260827-175312-a36a81

155,2 tok/s

NVIDIA GeForce RTX 3090 this runrun-20261101-195359-a93dda

145,7 tok/s

Hardware

GPU: NVIDIA GeForce RTX 3090 · 24 GB VRAM
CPU: AMD Ryzen 5 5600X 6-Core Processor
RAM: 31 GB
Mainboard: ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING

Setup

Runtime: tensorfold
Quantization: MLX-q4g64
Model: Qwen3.6-35B-A3B

Configuration

benchmark-konfiguration — run-20261101-195359-a93dda

# LLM-Benchmark Konfiguration # Modell : Qwen3.6-35B-A3B # Engine : tensorfold # Run-ID : run-20261101-195359-a93dda # GPU : NVIDIA GeForce RTX 3090 # CPU : AMD Ryzen 5 5600X 6-Core Processor # RAM : 31 GB bench@llm-benchmark:~$ tensorfold \ --model Qwen3.6-35B-A3B-MLX-4bit

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.	tensorfold
Kontextlaenge?Maximale Anzahl Tokens (Eingabe + erzeugte Ausgabe zusammen), die das Modell pro Anfrage verarbeiten kann.	8192
execution_type	local

All benchmarks of this model To leaderboard

Model comparison

Qwen3.6-35B-A3B 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 Workstation Edition 1.499,4 tok/sNVIDIA GeForce RTX 5090 1.409,4 tok/sNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 1.367,3 tok/sNVIDIA GeForce RTX 3090 Ti 852,0 tok/s★ NVIDIA GeForce RTX 3090 833,3 tok/s this runNVIDIA RTX A6000 800,4 tok/sAMD Radeon RX 7900 XTX 450,5 tok/sAMD Radeon PRO W7900 Dual Slot 296,2 tok/sAMD Radeon AI PRO R9700 271,4 tok/sAMD Radeon PRO W7800 48GB 235,5 tok/sNVIDIA Tesla V100 PCIe 32GB 225,9 tok/sNVIDIA DGX SPARK (GB10) 169,4 tok/sNVIDIA GeForce RTX 5070 Ti 142,9 tok/sNVIDIA RTX SPARK (N1X) 142,0 tok/sIntel Arc Pro B70 124,5 tok/sAMD Radeon 8060S Graphics 98,3 tok/sNVIDIA GeForce RTX 2060 27,5 tok/sCPU-only 4,3 tok/s

CPUby processor

★ AMD Ryzen 5 5600X 6-Core Processor 1.499,4 tok/s this runAMD Ryzen 9 9950X 16-Core Processor 1.489,1 tok/sAMD Ryzen 7 5800X3D 8-Core Processor 1.409,4 tok/sAMD Ryzen Threadripper PRO 9965WX 24-Cores 1.367,3 tok/sAMD Ryzen 9 8945HX with Radeon Graphics 852,0 tok/sAMD Ryzen Threadripper PRO 7955WX 16-Cores 800,4 tok/sAMD Ryzen Threadripper PRO 3955WX 16-Cores 450,5 tok/sAMD Ryzen Threadripper PRO 5975WX 32-Cores 296,2 tok/sAMD Ryzen 3 PRO 3200GE w/ Radeon Vega Graphics 225,9 tok/sAMD Ryzen 3 3100 4-Core Processor 198,8 tok/sIntel(R) Xeon(R) CPU E5-4657L v2 @ 2.40GHz 181,7 tok/sNVIDIA Grace 169,4 tok/sNVIDIA RTX Spark N1X (6144-core GPU, 20-core CPU) 142,0 tok/sAMD Ryzen 9 7945HX with Radeon Graphics 124,5 tok/sAMD RYZEN AI MAX+ 395 w/ Radeon 8060S 98,3 tok/sIntel(R) Xeon(R) CPU E5-4620 v2 @ 2.60GHz 4,3 tok/s

MBby mainboard

★ ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING 1.499,4 tok/s this runASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI 1.489,1 tok/sASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE 1.367,3 tok/sMeigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) 852,0 tok/sASUSTeK COMPUTER INC. PRIME A520M-K 830,5 tok/sASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 450,5 tok/sASUSTeK COMPUTER INC. ROG STRIX X570-F GAMING 225,9 tok/sDell Inc. PowerEdge R820 181,7 tok/sASUSTeK COMPUTER INC. GX10 169,4 tok/sASUS H7607BA 142,0 tok/sShenzhen Meigao Electronic Equipment Co.,Ltd DRFXI (MotherBoard Series) 124,5 tok/sBosgame AXB35-02 (BeyondMax Series) 98,3 tok/s

ENGby engine

llama.cpp 1.499,4 tok/svLLM 952,8 tok/s★ tensorfold 833,3 tok/s this rununbekannt 830,5 tok/s

DRVby driver

unbekannt 1.499,4 tok/sNVIDIA 590.48.01 / CUDA 13.1 169,4 tok/sIntel 26.18.38308.4 124,5 tok/sAMD 7.0.0-27-generic 79,4 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 12 W

⚡ TDP 362 W

☁️ External LLM (API) show

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

Electricity0.30 EUR/kWh

Avg power (incl. idle)362 W estimated (TDP)GPU 350 + Board 12 W full load

Avg cost / hourEUR 0.11

Electricity / 1M tokensEUR 0.21

Token / kWh1.45M

Acquisition (system)EUR 1,518 full priceGPU EUR 850 · CPU EUR 121 · Board EUR 149 · RAM EUR 248 · PSU EUR 150

Electricity (2 years)–

TCO (2 years)EUR 3,421

Output tokens (2 years)9.19B

☁️ 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 (12 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

	Qwen3.6-35B-A3BNVIDIA GeForce RTX 3090	Qwen3.6-35B-A3BNVIDIA RTX PRO 6000 Blackwell Workstation Edition	Qwen3.6-35B-A3B3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition	Qwen3.6-35B-A3B3x 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)	–	–	–	–

NVIDIA GeForce RTX 3090 · this run 146 tok/s NVIDIA RTX PRO 6000 Blackwell Workstation Edition 267 tok/s 3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 243 tok/s 3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 229 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.