Qwen2.5-72B-Instruct vs Llama-3.3-70B-Instruct
MeasuredOn the same M3 Ultra (same quant and runtime family, chat-short), Qwen2.5-72B-Instruct decoded at 16.3 tok/s and Llama-3.3-70B-Instruct at 16.8 tok/s, each model's fastest matched run on that hardware.
Signed, community-submitted decode tok/s, prefill, and TTFT for Qwen2.5-72B-Instruct vs Llama-3.3-70B-Instruct — every number links to the run it came from.
How to read these results
On the same M3 Ultra, Qwen2.5-72B-Instruct decoded at 16.3 tok/s and Llama-3.3-70B-Instruct at 16.8 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.
Hardware with data on both models
| Hardware | Qwen2.5-72B-Instruct decode | Llama-3.3-70B-Instruct decode | Δ | Source runs |
|---|---|---|---|---|
| M3 Ultra (60-core GPU) | 16.31tok/s chat-short · mlx · 4BIT (from file name) | 16.78tok/s chat-short · mlx · 4BIT (from file name) | -0.5 | r_5c80gthqlh6 · r_sx3a4y9n-m4 |
See also: Qwen2.5-72B-Instruct benchmarks · Llama-3.3-70B-Instruct benchmarks · All hardware · All models · Methodology