phi-4-4bit
7 workload results across 3 hardware configurations.
Fastest local config
140.9 decode tok/s
on RTX 5090 (32GB) + AMD Ryzen 7 9850X3D 8-Core Processor (8c) + 30GB via llama.cpp. see full run
For a standard 4-bit quantization using MLX, here are measured results for short-chat generation.
On MLX 4-bit, phi-4 is fastest on the M3 Ultra (60-core GPU) at 74.37 tok/s, 376% faster than the M3 Pro (18-core GPU) (15.61 tok/s).
| Card | Decode tok/s | TTFT (113-token prompt) | Run |
|---|---|---|---|
| M3 Ultra (60-core GPU) | 74.37 (single run) | 443 ms | r_sqzp-0rdez- |
| M3 Pro (18-core GPU) | 15.61 (single run) | 936 ms | r_-w8hnn61va_ |
phi-4, 4-bit, MLX, chat-short, batch 1. Only cards with their standard memory are listed; CPU-only hosts are excluded.
Local runs (7 workload results)
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 | - | 139.1tok/s | no data | 51.1ms | r_k1u_k1j_1i2 |
| chat-long | llama.cpp | - | 133.7tok/s | no data | 433ms | r_k1u_k1j_1i2 |
| concurrent-decode | llama.cpp | - | 140.2tok/s | no data | no data | r_k1u_k1j_1i2 |
| agent-trace | llama.cpp | - | 124.8tok/s | 21,162.4tok/s | 102ms | r_k1u_k1j_1i2 |
| chat-short | llama.cpp | - | 140.9tok/s | no data | 56.7ms | r_e-k4aea8ipr |
M3 Pro (18-core GPU) + 36GB unified
| Workload | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|
| chat-short | mlx@0.31.3 | - | 15.61tok/s | 120.7tok/s | 936ms | r_-w8hnn61va_ |
M3 Ultra (60-core GPU) + 96GB unified
| Workload | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|
| chat-short | mlx@0.31.3 | - | 74.37tok/s | 255.0tok/s | 443ms | r_sqzp-0rdez- |
phi-4-4bit on hardware
Request a phi-4-4bit run
No measured run yet for these pairs. Vote for the one you want; the most-requested runs go first. See all requests.
We never sell or share your email. Privacy