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

Qwen3.8-Flash-Next

Performance benchmark · measured on 30.09.2026 20:26

Benchmark-IDrun-20260930-183838-08e27e
Timebench 3 - Kombi (Prefill + Generation)MoE180BRuntime: llama.cppQuantisierung: UD-Q4_K_XL
Generation60,89tok/s
Prefill409,08tok/s
Time to First Token24.788,00ms
Total duration143,28s
Concurrency5parallel
Ranking in the field
319of 337 systems

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

This run is better than 5 % of all comparable systems.
Generation 60,9 tok/s
-75 % vs Ø 246,0
Prefill 409,1 tok/s
-96 % vs Ø 9.258,1
Time to First Token 24.788 ms
+554 % vs Ø 3.791
Distribution in the field19 – 909 tok/s
Ø 246 Median Dieser Lauf

Wie schlägt sich dieser Benchmark mit anderen Modellen?

How does this benchmark compare on other GPUs?

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

Hardware

GPU: 2x NVIDIA RTX A6000 · 48 GB VRAM
CPU: AMD EPYC 7203P 8-Core Processor
RAM: 252 GB
Mainboard: ASRockRack ROMED8-2T/BCM

Setup

Runtime: llama.cpp
Quantization: UD-Q4_K_XL
Model: Qwen3.8-Flash-Next

Configuration

benchmark-konfiguration — run-20260930-183838-08e27e
# LLM-Benchmark Konfiguration # Modell : Qwen3.8-Flash-Next # Engine : llama.cpp # Run-ID : run-20260930-183838-08e27e # GPU : 2x NVIDIA RTX A6000 # CPU : AMD EPYC 7203P 8-Core Processor # RAM : 252 GB bench@llm-benchmark:~$ llama-server Qwen3.8-Flash-Next-UD-Q4_K_XL \ -ngl 99 \ --n-cpu-moe 10 \ -fa on '(2xA6000' NVLink + 31GB RAM 'Offload)'
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
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.99
n-cpu-moe10
faon
execution_typelocal

All benchmarks of this model To leaderboard

Anzeige
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

89,567,445,223,10,9155192230268Prefill (tok/s)Generation (tok/s)AMD Radeon RX 7900 XTX - 18,7 tok/s Generation, 180 tok/s Prefill, TTFT 55.956 ms (3 Laufe)AMD Radeon RX 7900 XTXNVIDIA RTX A6000 - 71,7 tok/s Generation, 242 tok/s Prefill, TTFT 49.989 ms (9 Laufe) | DIESER LAUF★ NVIDIA RTX A6000
★ NVIDIA RTX A6000 71,7 tok/s this runAMD Radeon RX 7900 XTX 18,7 tok/s

CPUby processor

89,567,445,223,10,9155192230268Prefill (tok/s)Generation (tok/s)AMD Ryzen Threadripper PRO 3955WX 16-Cores - 18,7 tok/s Generation, 180 tok/s Prefill, TTFT 55.956 ms (3 Laufe)AMD Ryzen Threadrippe...AMD EPYC 7203P 8-Core Processor - 71,7 tok/s Generation, 242 tok/s Prefill, TTFT 49.989 ms (9 Laufe) | DIESER LAUF★ AMD EPYC 7203P 8-Core...
★ AMD EPYC 7203P 8-Core Processor 71,7 tok/s this runAMD Ryzen Threadripper PRO 3955WX 16-Cores 18,7 tok/s

MBby mainboard

86,269,953,637,321,0105224343461Prefill (tok/s)Generation (tok/s)ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI - 35,4 tok/s Generation, 170 tok/s Prefill, TTFT 59.933 ms (9 Laufe)ASUSTeK COMPUTER INC....ASRockRack ROMED8-2T/BCM - 71,7 tok/s Generation, 397 tok/s Prefill, TTFT 26.126 ms (3 Laufe) | DIESER LAUF★ ASRockRack ROMED8-2T/...
★ ASRockRack ROMED8-2T/BCM 71,7 tok/s this runASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 35,4 tok/s

ENGby engine

78,975,371,768,164,6213222231240Prefill (tok/s)Generation (tok/s)llama.cpp - 71,7 tok/s Generation, 227 tok/s Prefill, TTFT 51.481 ms (12 Laufe) | DIESER LAUF★ llama.cpp
★ llama.cpp 71,7 tok/s this run

DRVby driver

78,975,371,768,164,6213222231240Prefill (tok/s)Generation (tok/s)unbekannt - 71,7 tok/s Generation, 227 tok/s Prefill, TTFT 51.481 ms (12 Laufe)unbekannt
unbekannt 71,7 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 30 W
⚡ TDP 600 W
☁️ External LLM (API)
Electricity0.30 EUR/kWh
Avg power (incl. idle)600 W estimated (TDP)GPU 600 W full load
Avg cost / hourEUR 0.18
Electricity / 1M tokensEUR 0.82
Token / kWh365.34K
Acquisition (system)EUR 6,180 partial priceGPU EUR 6,000 · PSU EUR 180
Electricity (2 years)–
TCO (2 years)EUR 9,334
Output tokens (2 years)3.84B
☁️ 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 (30 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-Next2x NVIDIA RTX A6000Qwen3.8-Flash-Next2x NVIDIA RTX A6000Qwen3.8-Flash-Next2x NVIDIA RTX A6000Qwen3.8-Flash-Next2x AMD Radeon RX 7900 XTX
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.