Qwen2.5-Coder-32B-Instruct vs Codestral-22B-v0.1
MeasuredOn the same M3 Ultra (same quant and runtime family, chat-short), Qwen2.5-Coder-32B-Instruct decoded at 34.5 tok/s and Codestral-22B-v0.1 at 47.5 tok/s, each model's fastest matched run on that hardware.
Signed, community-submitted decode tok/s, prefill, and TTFT for Qwen2.5-Coder-32B-Instruct vs Codestral-22B-v0.1 — every number links to the run it came from.
How to read these results
On the same M3 Ultra, Qwen2.5-Coder-32B-Instruct decoded at 34.5 tok/s and Codestral-22B-v0.1 at 47.5 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 | Qwen2.5-Coder-32B-Instruct decode | Codestral-22B-v0.1 decode | Δ | Source runs |
|---|---|---|---|---|
| M3 Ultra (60-core GPU) | 34.48tok/s chat-short · mlx · 4BIT (from file name) | 47.49tok/s chat-short · mlx · 4BIT (from file name) | -13.0 | r_721b4bls_oq · r_79dvtag5fd_ |
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 | Qwen2.5-Coder-32B-Instruct decode | Codestral-22B-v0.1 decode | Why it isn't matched | Runs |
|---|---|---|---|---|
| RTX 5090 (32GB) | 71.91tok/s llama.cpp · quant unreported | 93.72tok/s llama.cpp · quant unreported | quant unreported | r_nkbs6d3-d21 · r_4q040m4scic |
See also: Qwen2.5-Coder-32B-Instruct benchmarks · Codestral-22B-v0.1 benchmarks · All hardware · All models · Methodology