Qwen2.5-7B-Instruct-4bit
8 workload results across 3 hardware configurations.
Fastest local config
246.7 decode tok/s
on RTX 5090 (32GB) + AMD Ryzen 7 9850X3D 8-Core Processor (8c) + 30GB via llama.cpp. see full run
Local runs (8 runs)
Runs from contributors' own machines via MLX, llama.cpp, vLLM, exllamav2, or ollama. Signed on the submitter's hardware.
RTX 5090 (32GB) + AMD Ryzen 7 9850X3D 8-Core Processor (8c) + 30GB
| Workload | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|
| chat-short | llama.cpp | - | 232.7tok/s | no data | 37.0ms | r_1shiviswt3d |
| chat-long | llama.cpp | - | 232.4tok/s | no data | 239ms | r_1shiviswt3d |
| concurrent-decode | llama.cpp | - | 240.7tok/s | no data | no data | r_1shiviswt3d |
| agent-trace | llama.cpp | - | 232.5tok/s | 38,007.7tok/s | 57.1ms | r_1shiviswt3d |
| chat-short | llama.cpp | - | 246.7tok/s | no data | 33.9ms | r_3yn-4321hp- |
M3 Pro (18-core GPU) + 36GB unified
| Workload | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|
| chat-short | mlx@0.31.3 | 4bit | 30.52tok/s | 161.9tok/s | 809ms | r_llzv_g-ymaf |
| chat-short | mlx@0.31.3 | 4bit | 15.67tok/s | 54.34tok/s | 2,411ms | r_e3t93rscswq |
M3 Ultra (60-core GPU) + 96GB unified
| Workload | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|
| chat-short | mlx@0.31.3 | 4bit | 139.6tok/s | 190.0tok/s | 689ms | r_5r6rhiynenc |