Qwen2.5-Coder-14B-Instruct vs Qwen3-Coder-30B-A3B-Instruct
MeasuredOn the same RTX 5090 (same quant and runtime family, chat-short), Qwen2.5-Coder-14B-Instruct decoded at 137.3 tok/s and Qwen3-Coder-30B-A3B-Instruct at 304.9 tok/s, each model's fastest matched run on that hardware.
Signed, community-submitted decode tok/s, prefill, and TTFT for Qwen2.5-Coder-14B-Instruct vs Qwen3-Coder-30B-A3B-Instruct — every number links to the run it came from.
How to read these results
On the same RTX 5090, Qwen2.5-Coder-14B-Instruct decoded at 137.3 tok/s and Qwen3-Coder-30B-A3B-Instruct at 304.9 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-Coder-14B-Instruct decode | Qwen3-Coder-30B-A3B-Instruct decode | Δ | Source runs |
|---|---|---|---|---|
| RTX 5090 (32GB) | 137.3tok/s chat-short · llama.cpp · Q4_K_M | 304.9tok/s chat-short · llama.cpp · Q4_K_M | -167.6 | r_xnr_k4emoyr · r_vsype4pm1ko |
| RTX 4090 (24GB) | 89.87tok/s chat-short · llama.cpp · Q4_K_M | 232.5tok/s chat-short · llama.cpp · Q4_K_M | -142.6 | r_eqnsqmgz29a · r_1k8942tw3zf |
| RTX 3090 (24GB) | 78.21tok/s chat-short · llama.cpp · Q4_K_M | 199.7tok/s chat-short · llama.cpp · Q4_K_M | -121.5 | r_rkyf9wxzw9z · r_0dh1kzbv4h0 |
| M3 Ultra (60-core GPU) | 70.51tok/s chat-short · mlx · 4BIT (from file name) | 112.2tok/s chat-short · mlx · 4BIT (from file name) | -41.6 | r_l36cijqxq4t · r_fpsca03u2o_ |
See also: Qwen2.5-Coder-14B-Instruct benchmarks · Qwen3-Coder-30B-A3B-Instruct benchmarks · All hardware · All models · Methodology