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

Devstral-2-123B-Instruct-2512

Performance benchmark · measured on 13.08.2026 09:32

Benchmark-IDrun-20260813-073605-793dd1
Timebench 3 - Kombi (Prefill + Generation)Dense123BRuntime: llama.cpp batched-bench ROCmQuantisierung: Q4_K_M
Generation7,19tok/s
Prefill276,37tok/s
Time to First Token1.853,00ms
Total duration0,00s
Concurrency1parallel
Ranking in the field
362of 455 systems

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

This run is better than 20 % of all comparable systems.
Generation 7,2 tok/s
-80 % vs Ø 35,3
Prefill 276,4 tok/s
-86 % vs Ø 1.966,4
Time to First Token 1.853 ms
-95 % vs Ø 35.848
Distribution in the field1 – 148 tok/s
Ø 35 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: 4x AMD Radeon AI PRO R9700 · 32 GB VRAM
CPU: AMD Ryzen Threadripper PRO 7955WX 16-Cores
RAM: 184 GB
Mainboard: ASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE

Setup

Runtime: llama.cpp batched-bench ROCm
Quantization: Q4_K_M
Model: Devstral-2-123B-Instruct-2512

Anmerkung

llama-batched-bench, 4x AMD Radeon AI PRO R9700 [4 verbaut, 4 genutzt], voll GPU (kein RAM-Offload), GPU-Nutzung verifiziert, npl=1 (pp512/tg128), ngl=99

Configuration

benchmark-konfiguration — run-20260813-073605-793dd1
# LLM-Benchmark Konfiguration # Modell : Devstral-2-123B-Instruct-2512 # Engine : llama.cpp # Run-ID : run-20260813-073605-793dd1 # GPU : 4x AMD Radeon AI PRO R9700 # CPU : AMD Ryzen Threadripper PRO 7955WX 16-Cores # RAM : 184 GB bench@llm-benchmark:~$ llama-batched-bench (llama.cpp ROCm) \ --model /home/godcore/.cache/huggingface/hub/models--lmstudio-community--Devstral-2-123B-Instruct-2512-GGUF/snapshots/330d61a4b7ed244a64a6bc3f7e98b250c3264546/Devstral-2-123B-Instruct-2512-Q4_K_M-00001-of-00002.gguf \ -ngl 99 \ -c 8192 \ -b 2048 \ -ub 512 \ -npp 512 \ -ntg 128 \ -npl 1 \ HIP_VISIBLE_DEVICES 0,1,2,3
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.llama-batched-bench (llama.cpp ROCm)
Modellalias?Der Name, unter dem das Modell ueber die API angesprochen wird. Genau dieser Wert muss im Request-Feld 'model' stehen.Devstral-2-123B-Instruct-2512
Kontextlaenge?Maximale Anzahl Tokens (Eingabe + erzeugte Ausgabe zusammen), die das Modell pro Anfrage verarbeiten kann.8192
Modellpfad?Pfad zur GGUF-Modelldatei, die geladen und ausgeliefert wird./home/godcore/.cache/huggingface/hub/models--lmstudio-community--Devstral-2-123B-Instruct-2512-GGUF/snapshots/330d61a4b7ed244a64a6bc3f7e98b250c3264546/Devstral-2-123B-Instruct-2512-Q4_K_M-00001-of-00002.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.99
Kontext?Groesse des Kontextfensters in Token. 0 = der beim Training verwendete Kontext des Modells.8192
Batch-Groesse?Logische Batchgroesse fuer die Prompt-Verarbeitung (Anzahl Token pro Prefill-Durchlauf).2048
Micro-Batch?Physische Micro-Batchgroesse. Begrenzt, wie viele Token gleichzeitig durch die Compute-Pipeline laufen.512
npp512
ntg128
npl1
HIP_VISIBLE_DEVICES0,1,2,3

All benchmarks of this model To leaderboard

Anzeige
Model comparison

Devstral-2-123B-Instruct-2512 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

18613992,846,40,005731.1471.720Prefill (tok/s)Generation (tok/s)NVIDIA RTX PRO 6000 Blackwell Workstation Edition - 143,1 tok/s Generation, 1.403 tok/s Prefill, TTFT 31.598 ms (3 Laufe)NVIDIA RTX PRO 6000 B...NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition - 123,5 tok/s Generation, 1.389 tok/s Prefill, TTFT 39.803 ms (9 Laufe)NVIDIA RTX PRO 6000 B...NVIDIA GeForce RTX 5070 Ti - 4,5 tok/s Generation, 127 tok/s Prefill, TTFT 54.082 ms (3 Laufe)NVIDIA GeForce RTX 50...NVIDIA GeForce RTX 3090 Ti - 2,9 tok/s Generation, 129 tok/s Prefill, TTFT 54.597 ms (3 Laufe)NVIDIA GeForce RTX 30...NVIDIA GeForce RTX 5090 - 2,1 tok/s Generation, 113 tok/s Prefill, TTFT 62.542 ms (3 Laufe)NVIDIA GeForce RTX 50...AMD Radeon AI PRO R9700 - 50,0 tok/s Generation, 335 tok/s Prefill, TTFT 137.442 ms (29 Laufe) | DIESER LAUF★ AMD Radeon AI PRO R97...
NVIDIA RTX PRO 6000 Blackwell Workstation Edition 143,1 tok/sNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 123,5 tok/s★ AMD Radeon AI PRO R9700 50,0 tok/s this runNVIDIA GeForce RTX 5070 Ti 4,5 tok/sNVIDIA GeForce RTX 3090 Ti 2,9 tok/sNVIDIA GeForce RTX 5090 2,1 tok/s

CPUby processor

18613992,846,40,005731.1471.720Prefill (tok/s)Generation (tok/s)AMD Ryzen 9 9950X 16-Core Processor - 143,1 tok/s Generation, 1.403 tok/s Prefill, TTFT 31.598 ms (3 Laufe)AMD Ryzen 9 9950X 16-...AMD Ryzen Threadripper PRO 9965WX 24-Cores - 123,5 tok/s Generation, 1.389 tok/s Prefill, TTFT 39.803 ms (9 Laufe)AMD Ryzen Threadrippe...AMD Ryzen Threadripper PRO 5975WX 32-Cores - 4,5 tok/s Generation, 127 tok/s Prefill, TTFT 54.082 ms (3 Laufe)AMD Ryzen Threadrippe...AMD Ryzen 9 8945HX with Radeon Graphics - 2,9 tok/s Generation, 129 tok/s Prefill, TTFT 54.597 ms (3 Laufe)AMD Ryzen 9 8945HX wi...AMD Ryzen 7 5800X3D 8-Core Processor - 2,1 tok/s Generation, 113 tok/s Prefill, TTFT 62.542 ms (3 Laufe)AMD Ryzen 7 5800X3D 8...AMD Ryzen Threadripper PRO 7955WX 16-Cores - 50,0 tok/s Generation, 335 tok/s Prefill, TTFT 137.442 ms (29 Laufe) | DIESER LAUF★ AMD Ryzen Threadrippe...
AMD Ryzen 9 9950X 16-Core Processor 143,1 tok/sAMD Ryzen Threadripper PRO 9965WX 24-Cores 123,5 tok/s★ AMD Ryzen Threadripper PRO 7955WX 16-Cores 50,0 tok/s this runAMD Ryzen Threadripper PRO 5975WX 32-Cores 4,5 tok/sAMD Ryzen 9 8945HX with Radeon Graphics 2,9 tok/sAMD Ryzen 7 5800X3D 8-Core Processor 2,1 tok/s

MBby mainboard

18613992,846,40,005731.1471.720Prefill (tok/s)Generation (tok/s)ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI - 143,1 tok/s Generation, 1.403 tok/s Prefill, TTFT 31.598 ms (3 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI - 4,5 tok/s Generation, 127 tok/s Prefill, TTFT 54.082 ms (3 Laufe)ASUSTeK COMPUTER INC....Meigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) - 2,9 tok/s Generation, 129 tok/s Prefill, TTFT 54.597 ms (3 Laufe)Meigao Innovation Tec...ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING - 2,1 tok/s Generation, 113 tok/s Prefill, TTFT 62.542 ms (3 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE - 123,5 tok/s Generation, 585 tok/s Prefill, TTFT 114.317 ms (38 Laufe) | DIESER LAUF★ ASUSTeK COMPUTER INC....
ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI 143,1 tok/s★ ASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE 123,5 tok/s this runASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 4,5 tok/sMeigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) 2,9 tok/sASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING 2,1 tok/s

ENGby engine

18513892,346,20,0248376503631Prefill (tok/s)Generation (tok/s)unbekannt - 6,5 tok/s Generation, 323 tok/s Prefill, TTFT 6.557 ms (1 Lauf)unbekanntllama.cpp - 143,1 tok/s Generation, 556 tok/s Prefill, TTFT 100.938 ms (49 Laufe) | DIESER LAUF★ llama.cpp
★ llama.cpp 143,1 tok/s this rununbekannt 6,5 tok/s

DRVby driver

18513892,346,20,0248376503631Prefill (tok/s)Generation (tok/s)unbekannt - 143,1 tok/s Generation, 556 tok/s Prefill, TTFT 100.938 ms (49 Laufe)unbekanntAMD 7.0.0-27-generic - 6,5 tok/s Generation, 323 tok/s Prefill, TTFT 6.557 ms (1 Lauf)AMD 7.0.0-27-generic
unbekannt 143,1 tok/sAMD 7.0.0-27-generic 6,5 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 85 W
⚡ TDP 371 W
☁️ External LLM (API)
Electricity0.30 EUR/kWh
Avg power (incl. idle)371 W estimated (TDP)GPU 300 + CPU 46 + Board 25 W full load
Avg cost / hourEUR 0.11
Electricity / 1M tokensEUR 4.30
Token / kWh69.77K
Acquisition (system)EUR 6,824 full priceGPU EUR 1,400 · CPU EUR 1,399 · Board EUR 1,299 · RAM EUR 2,576 · PSU EUR 150
Electricity (2 years)
TCO (2 years)EUR 8,774
Output tokens (2 years)453.49M
☁️ 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 (85 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

Devstral-2-123B-Instruct-2512AMD Radeon AI PRO R9700Devstral-2-123B-Instruct-2512NVIDIA RTX PRO 6000 Blackwell Workstation EditionDevstral-2-123B-Instruct-25123x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionDevstral-2-123B-Instruct-25123x 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.