Created bymario-alka.dePowered bygodcore.denoob2claw.detricoma.de
Contributed byMario AlkaQwen (Alibaba)

Qwen3-4B

Performance benchmark · measured on 27.09.2026 00:21

Benchmark-IDrun-20260926-225053-6edbc5
Timebench 3 - Kombi (Prefill + Generation)Dense4BRuntime: llama.cppQuantisierung: Q4_K_M
Generation211,36tok/s
Prefill8.014,23tok/s
Time to First Token1.553,50ms
Total duration51,79s
Concurrency5parallel
Ranking in the field
567of 1352 systems

Performance benchmark · Primary metric: Generation-Speed (tok/s) · 5× concurrent

This run is better than 58 % of all comparable systems.
Generation 211,4 tok/s
-17 % vs Ø 253,9
Prefill 8.014,2 tok/s
+47 % vs Ø 5.462,2
Time to First Token 1.554 ms
-95 % vs Ø 28.760
Distribution in the field0 – 1.349 tok/s
Ø 254 Median Dieser Lauf

Wie schlägt sich dieser Benchmark mit anderen Modellen?

gemma-4-E2B-itNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260727-181606-0adf6d
1.349,3 tok/s
gemma-4-E2B-itNVIDIA GeForce RTX 5090 · run-20260728-184455-5b4937
1.301,9 tok/s
gemma-4-E2B-it3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-184456-4111a9
1.195,7 tok/s
Nemotron-3-Nano-4BNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260727-181607-30d864
1.188,7 tok/s
Nemotron-3-Nano-4BNVIDIA GeForce RTX 5090 · run-20260728-194135-5432cd
1.164,8 tok/s
gpt-oss-20bNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260728-140954-9e81f9
1.136,3 tok/s
gpt-oss-20bNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260730-035052-09627f
1.135,0 tok/s
gpt-oss-20bNVIDIA GeForce RTX 5090 · run-20260729-032121-058a31
1.113,5 tok/s
Nemotron-3-Nano-4B3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-194136-836a91
1.090,6 tok/s
Nemotron-3-Nano-4B3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260727-181606-8b381e
1.077,4 tok/s
Nemotron-3-Nano-4B3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-194136-b8a818
1.071,0 tok/s
Nemotron-3-Nano-4B3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-194136-038e92
1.051,1 tok/s
gpt-oss-20b3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260729-032121-ffb18a
1.015,5 tok/s
Nemotron-3-Nano-30B-A3BNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260729-032121-305ad6
1.007,8 tok/s
Qwen3-4B this run2× AMD Radeon RX 7900 XTX · run-20260926-225053-6edbc5
211,4 tok/s

How does this benchmark compare on other GPUs?

Same model on different hardware · 5× concurrent · Generation (tok/s)

Hardware

GPU: 2x AMD Radeon RX 7900 XTX · 24 GB VRAM
CPU: AMD Ryzen Threadripper PRO 3955WX 16-Cores
RAM: 63 GB
Mainboard: ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI

Setup

Runtime: llama.cpp
Quantization: Q4_K_M
Model: Qwen3-4B

Configuration

benchmark-konfiguration — run-20260926-225053-6edbc5
# LLM-Benchmark Konfiguration # Modell : Qwen3-4B # Engine : llama.cpp # Run-ID : run-20260926-225053-6edbc5 # GPU : 2x AMD Radeon RX 7900 XTX # CPU : AMD Ryzen Threadripper PRO 3955WX 16-Cores # RAM : 63 GB bench@llm-benchmark:~$ llama-server \ -m Qwen3-4B-Q4_K_M.gguf \ --alias Qwen3-4B \ -ngl 999 \ -fa on \ -c 40960 \ -np 10 '(2x' RX 7900 XTX 'ROCm)'
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.llamacpp
Modellalias?Der Name, unter dem das Modell ueber die API angesprochen wird. Genau dieser Wert muss im Request-Feld 'model' stehen.Qwen3-4B
Kontextlaenge?Maximale Anzahl Tokens (Eingabe + erzeugte Ausgabe zusammen), die das Modell pro Anfrage verarbeiten kann.40960
Modellpfad?Pfad zur GGUF-Modelldatei, die geladen und ausgeliefert wird.Qwen3-4B-Q4_K_M.gguf
GPU-Layer?Anzahl der auf die GPU ausgelagerten Modell-Layer. Hoeher = mehr VRAM und schneller; der Rest laeuft auf der CPU. 999 = alles auf GPU.999
faon
Kontext?Groesse des Kontextfensters in Token. 0 = der beim Training verwendete Kontext des Modells.40960
np10

All benchmarks of this model To leaderboard

Anzeige
Model comparison

Qwen3-4B 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

3593413233052875.8558.00610.15712.308Prefill (tok/s)Generation (tok/s)AMD Radeon AI PRO R9700 - 325,2 tok/s Generation, 7.193 tok/s Prefill, TTFT 1.751 ms (3 Laufe)AMD Radeon AI PRO R97...AMD Radeon RX 7900 XTX - 320,4 tok/s Generation, 10.970 tok/s Prefill, TTFT 1.493 ms (6 Laufe) | DIESER LAUF★ AMD Radeon RX 7900 XTX
AMD Radeon AI PRO R9700 325,2 tok/s★ AMD Radeon RX 7900 XTX 320,4 tok/s this run

CPUby processor

3593413233052875.8558.00610.15712.308Prefill (tok/s)Generation (tok/s)AMD Ryzen 3 3100 4-Core Processor - 325,2 tok/s Generation, 7.193 tok/s Prefill, TTFT 1.751 ms (3 Laufe)AMD Ryzen 3 3100 4-Co...AMD Ryzen Threadripper PRO 3955WX 16-Cores - 320,4 tok/s Generation, 10.970 tok/s Prefill, TTFT 1.493 ms (6 Laufe) | DIESER LAUF★ AMD Ryzen Threadrippe...
AMD Ryzen 3 3100 4-Core Processor 325,2 tok/s★ AMD Ryzen Threadripper PRO 3955WX 16-Cores 320,4 tok/s this run

MBby mainboard

3593413233052875.8558.00610.15712.308Prefill (tok/s)Generation (tok/s)ASUSTeK COMPUTER INC. ROG STRIX X570-F GAMING - 325,2 tok/s Generation, 7.193 tok/s Prefill, TTFT 1.751 ms (3 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI - 320,4 tok/s Generation, 10.970 tok/s Prefill, TTFT 1.493 ms (6 Laufe) | DIESER LAUF★ ASUSTeK COMPUTER INC....
ASUSTeK COMPUTER INC. ROG STRIX X570-F GAMING 325,2 tok/s★ ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 320,4 tok/s this run

ENGby engine

3593413233052875.5519.04312.53416.026Prefill (tok/s)Generation (tok/s)vLLM - 320,4 tok/s Generation, 14.021 tok/s Prefill, TTFT 1.458 ms (3 Laufe)vLLMllama.cpp - 325,2 tok/s Generation, 7.556 tok/s Prefill, TTFT 1.640 ms (6 Laufe) | DIESER LAUF★ llama.cpp
★ llama.cpp 325,2 tok/s this runvLLM 320,4 tok/s

DRVby driver

3583423253092939.1289.5179.90510.294Prefill (tok/s)Generation (tok/s)unbekannt - 325,2 tok/s Generation, 9.711 tok/s Prefill, TTFT 1.579 ms (9 Laufe)unbekannt
unbekannt 325,2 tok/s
💰 Economics

Economics of this run

Operating cost, TCO and comparison with the next-best runs of the same model at identical concurrency (5× 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 10 W
⚡ TDP 720 W
☁️ External LLM (API)
Electricity0.30 EUR/kWh
Avg power (incl. idle)720 W estimated (TDP)GPU 710 + Board 10 W full load
Avg cost / hourEUR 0.22
Electricity / 1M tokensEUR 0.28
Token / kWh1.06M
Acquisition (system)EUR 2,782 partial priceGPU EUR 2,098 · RAM EUR 504 · PSU EUR 180
Electricity (2 years)–
TCO (2 years)EUR 6,566
Output tokens (2 years)13.33B
☁️ 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 (10 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

Qwen3-4B2x AMD Radeon RX 7900 XTXQwen3-4B2x AMD Radeon RX 7900 XTXQwen3-4BAMD Radeon AI PRO R9700
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.