Vergleich

Lokale KI-Hardware mit workloadbezogener Auswertung vergleichen.

Hardware A

Hardware BWorkloadOperating mode

Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.

Operating mode: Balanced. Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.

RTX 4090 24GB gewinnt for coding in balanced mode

Basierend auf Modelleignung, Geschwindigkeit und Qualität der Top-Empfehlungen.

SQwen 3.5 9B

llama.cppq4-k-mRuns well

9.8 GB / 12.0 GB

72.0 tok/s32K ctx

ACodeGeeX 4 9B

llama.cppq4-k-mRuns well

8.2 GB / 12.0 GB

75.3 tok/s116K ctx

AGemma 4 E4B

llama.cppq4-k-mRuns well

8.3 GB / 12.0 GB

63.1 tok/s63K ctx

Gewinner

SCodestral 2 25.08

llama.cppq4-k-mRuns well

19.2 GB / 24.0 GB

42.0 tok/s48K ctx

SQwen 3.6 27B

llama.cppq4-k-mTight fit

20.7 GB / 24.0 GB

31.7 tok/s69K ctx

SDevstral Small 2 24B Instruct

llama.cppq4-k-mTight fit

20.4 GB / 24.0 GB

40.0 tok/s40K ctx

Schnellvergleich