Leaderboard/models/qwen3-coder-30b-a3b-instruct

Qwen3-Coder-30B-A3B-Instruct-4bit

18 workload results across 5 hardware configurations.

Fastest local config

304.9 decode tok/s

on RTX 5090 (32GB) + AMD EPYC 7543 32-Core Processor (120c) + 945GB via llama.cpp (Q4_K_M). see full run

For a standard Q4_K_M quantization using llama.cpp, here are measured results for short-chat generation.

On llama.cpp Q4_K_M, Qwen3-Coder-30B-A3B-Instruct is fastest on the RTX 5090 at 304.9 tok/s, 31% faster than the RTX 4090 (232.5 tok/s); slowest: RTX 3090 at 199.7 tok/s.

Card	Decode tok/s	TTFT (110-token prompt)	Run
RTX 5090	304.9 (single run)	214 ms	r_vsype4pm1ko
RTX 4090	232.5 (single run)	132 ms	r_1k8942tw3zf
RTX 3090	199.7 (single run)	180 ms	r_0dh1kzbv4h0

Qwen3-Coder-30B-A3B-Instruct, Q4_K_M, llama.cpp, chat-short, batch 1. Only cards with their standard memory are listed; CPU-only hosts are excluded. Runs use different llama.cpp versions (RTX 5090: 0.7.0-dev (build 1, commit 1268eda); RTX 4090: 0.6.0-dev (build 1, commit df18bee); RTX 3090: 0.7.0-dev (build 1, commit ea0c4c2)), so treat this as measured reference points, not a controlled ranking.

Local runs (18 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 EPYC 7543 32-Core Processor (120c) + 945GB

Workload	Backend	Quant	decode tok/s	prefill tok/s	TTFT	Run
chat-short	llama.cpp@0.7.0-dev (build 1, commit 1268eda)	Q4_K_M	304.9tok/s	514.5tok/s	214ms	r_vsype4pm1ko
chat-long	llama.cpp@0.7.0-dev (build 1, commit 1268eda)	Q4_K_M	266.2tok/s	8,353.2tok/s	377ms	r_vsype4pm1ko
concurrent-decode	llama.cpp@0.7.0-dev (build 1, commit 1268eda)	Q4_K_M	296.7tok/s	no data	no data	r_vsype4pm1ko
agent-trace	llama.cpp@0.7.0-dev (build 1, commit 1268eda)	Q4_K_M	278.8tok/s	23,318.7tok/s	96.6ms	r_vsype4pm1ko

RTX 3090 (24GB) + AMD EPYC 7282 16-Core Processor (16c) + 126GB

Workload	Backend	Quant	decode tok/s	prefill tok/s	TTFT	Run
chat-short	llama.cpp@0.7.0-dev (build 1, commit ea0c4c2)	Q4_K_M	199.7tok/s	612.5tok/s	180ms	r_0dh1kzbv4h0
chat-long	llama.cpp@0.7.0-dev (build 1, commit ea0c4c2)	Q4_K_M	187.0tok/s	3,971.3tok/s	792ms	r_0dh1kzbv4h0
concurrent-decode	llama.cpp@0.7.0-dev (build 1, commit ea0c4c2)	Q4_K_M	198.0tok/s	no data	no data	r_0dh1kzbv4h0
agent-trace	llama.cpp@0.7.0-dev (build 1, commit ea0c4c2)	Q4_K_M	190.9tok/s	10,182.2tok/s	214ms	r_0dh1kzbv4h0

RTX 4090 (24GB) + AMD EPYC 7642 48-Core Processor (48c) + 252GB

Workload	Backend	Quant	decode tok/s	prefill tok/s	TTFT	Run
chat-short	llama.cpp@0.6.0-dev (build 1, commit df18bee)	Q4_K_M	232.5tok/s	831.9tok/s	132ms	r_1k8942tw3zf
chat-long	llama.cpp@0.6.0-dev (build 1, commit df18bee)	Q4_K_M	206.9tok/s	7,688.7tok/s	409ms	r_1k8942tw3zf
concurrent-decode	llama.cpp@0.6.0-dev (build 1, commit df18bee)	Q4_K_M	225.1tok/s	no data	no data	r_1k8942tw3zf
agent-trace	llama.cpp@0.6.0-dev (build 1, commit df18bee)	Q4_K_M	216.4tok/s	19,138.1tok/s	108ms	r_1k8942tw3zf

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	-	68.41tok/s	no data	334ms	r_pm_a1uf2ufc
chat-long	llama.cpp	-	70.55tok/s	no data	1,204ms	r_pm_a1uf2ufc
concurrent-decode	llama.cpp	-	68.04tok/s	no data	no data	r_pm_a1uf2ufc
agent-trace	llama.cpp	-	66.09tok/s	4,816.7tok/s	451ms	r_pm_a1uf2ufc
chat-short	llama.cpp	-	259.9tok/s	no data	218ms	r_c7qyvvmmsv1

M3 Ultra (60-core GPU) + 96GB unified

Workload	Backend	Quant	decode tok/s	prefill tok/s	TTFT	Run
chat-short	mlx@0.31.3	-	112.2tok/s	204.0tok/s	539ms	r_fpsca03u2o_