Real performance
TTFT, prefill and generation tokens/s instead of theoretical peak numbers.
We benchmark AI language models on real hardware: speed in tokens per second and response time – plus quality on real agent tasks. Transparent, reproducible, directly comparable.
TTFT, prefill and generation tokens/s instead of theoretical peak numbers.
Points and success rates from complete agent and chat runs.
Model, quantization, runtime and hardware stay visible on every run.
Speed (tokens/s) and quality (harness) of the most-tested models – directly comparable.
| # | Model / Maker | Metrics | Parallel | GPU / CPU / RAM | Runtime | ||
|---|---|---|---|---|---|---|---|
| 1 | NVIDIA-Nemotron-3.5-Lightning-30B-A3B31.6BNVIDIA Performance benchmarkTimebench 3 - Kombi (Prefill + Generation) | 3.484,32 tok/s TG Prefill 8.256 · TTFT 14.388 ms | 10× | NVIDIA GeForce RTX 3090AMD Ryzen 5 5600X 6-Core Processor · 31 GB RAM | tensorfoldgodclaw 1.1.6MLX 4-bit | Details → | |
| 2 | NVIDIA-Nemotron-3.5-Lightning-30B-A3B31.6BNVIDIA Performance benchmarkTimebench 3 - Kombi (Prefill + Generation) | 2.867,15 tok/s TG Prefill 8.107 · TTFT 17.054 ms | 10× | NVIDIA GeForce RTX 3090 iAMD Ryzen 5 5600X 6-Core Processor · 31 GB RAM | tensorfoldgodclaw 1.1.6MLX 4-bit | Details → | |
| 3 | gemma-4-E2B-it5BGoogle Performance benchmarkTimebench 3 - Kombi (Prefill + Generation) | 2.491,20 tok/s TG Prefill 15.814 · TTFT 4.382 ms | 10× | NVIDIA GeForce RTX 5090AMD Ryzen 7 5800X3D 8-Core Processor · 126 GB RAM | llama.cppcodex_cliQ4_K_M | Details → | |
| 4 | gemma-4-E2B-it5BGoogle Performance benchmarkTimebench 3 - Kombi (Prefill + Generation) | 2.414,37 tok/s TG Prefill 20.277 · TTFT 4.107 ms | 10× | NVIDIA RTX PRO 6000 Blackwell Workstation EditionAMD Ryzen 9 9950X 16-Core Processor · 92 GB RAM | llama.cppgodclawQ4_K_M | Details → | |
| 5 | gemma-4-E2B-it5BGoogle Performance benchmarkTimebench 3 - Kombi (Prefill + Generation) | 2.412,27 tok/s TG Prefill 19.677 · TTFT 4.107 ms | 10× | NVIDIA RTX PRO 6000 Blackwell Workstation EditionAMD Ryzen 5 5600X 6-Core Processor · 31 GB RAM | llama.cppgodclawQ4_K_M | Details → |
Jede Kachel ist ein gemessener Lauf - Motiv, Modell und Generierungszeit auf echter Hardware.
Weitwinkel: Cyberpunk-Rechenzentrum
Qwen-Image-2.1
Detail: GPU-Rack Nahaufnahme
Qwen-Image-2.1
Figur: Technikerin im Serverraum
Qwen-Image-2.1
Textrendering: Server-Display
Qwen-Image-2.1
Weitwinkel: Cyberpunk-Rechenzentrum
Z-Image
Detail: GPU-Rack Nahaufnahme
Z-Image
Figur: Technikerin im Serverraum
Z-Image
Textrendering: Server-Display
Z-Image
Generation throughput (tokens/s) per GPU – minimum, average and maximum across all performance runs.
Generation throughput (tokens/s) per CPU – minimum, average and maximum across all performance runs.
We use technically necessary cookies. External content (e.g. YouTube videos) is only loaded with your consent and then sets cookies from YouTube/Google. More in our privacy policy.