Phi-4 vs Gemma-2-9b-it
MeasuredOn the same M3 Ultra (same quant and runtime family, chat-short), Phi-4 decoded at 74.4 tok/s and Gemma-2-9b-it at 89.5 tok/s, each model's fastest matched run on that hardware.
Signed, community-submitted decode tok/s, prefill, and TTFT for Phi-4 vs Gemma-2-9b-it — every number links to the run it came from.
How to read these results
On the same M3 Ultra, Phi-4 decoded at 74.4 tok/s and Gemma-2-9b-it at 89.5 tok/s. Each row in the table is a like-for-like pair: the same hardware, quant and runtime family on the chat-short workload, using each side's fastest such run. Runs can still differ in host, runtime build and context. These are reference results, not a controlled speed or answer-quality ranking. Inspect both source runs before choosing hardware. 1 more hardware config was measured on both but not like-for-like; they are listed separately in the "Not a like-for-like pair" table below and left out of the comparison.
Hardware with data on both models
| Hardware | Phi-4 decode | Gemma-2-9b-it decode | Δ | Source runs |
|---|---|---|---|---|
| M3 Ultra (60-core GPU) | 74.37tok/s chat-short · mlx · 4BIT (from file name) | 89.45tok/s chat-short · mlx · 4BIT (from file name) | -15.1 | r_sqzp-0rdez- · r_iz137eqvuzy |
| M3 Pro (18-core GPU) | 15.61tok/s chat-short · mlx · 4BIT (from file name) | 21.97tok/s chat-short · mlx · 4BIT (from file name) | -6.4 | r_-w8hnn61va_ · r_q7t7b7dcuz5 |
Not a like-for-like pair
These have data on both sides, but the model size, quant or runtime differs, so they are not a head-to-head and are left out of the comparison above.
| Hardware | Phi-4 decode | Gemma-2-9b-it decode | Why it isn't matched | Runs |
|---|---|---|---|---|
| RTX 5090 (32GB) | 140.9tok/s llama.cpp · quant unreported | 152.7tok/s llama.cpp · quant unreported | quant unreported | r_e-k4aea8ipr · r__b_bzmmab_8 |
See also: Phi-4 benchmarks · Gemma-2-9b-it benchmarks · All hardware · All models · Methodology