ModelDock / hardware record

RTX 3090 local AI benchmarks

Compare the published model-specific reading / prefill and single-stream writing / decode results for RTX 3090.

Memory24 GBPublished model records: 2

Available model records

RTX 3090 performance by model

ModelReading / prefillWriting / decodeEvidenceAs of

Qwen3.8 27Bqwen38_27b_c1_personal

705 tok/s

55.9 tok/s

Source-backed

2026-08-23

Qwen3.8 Flash Nextqwen38_flash_next_c1_personal

~116 tok/s

5.8 tok/s

Partial estimate

2026-08-28

Reading / prefill

Prefill · tok/s

Writing / decode

Decode · single stream · tok/s

Evidence

Each row is one hardware × model record. Source observations stay on Evidence; projections link to their calibration details on Data.