RTX 5090 32GB, llama.cpp CUDA: Qwen3-32B Q4 decode tok/s @4K ctx (Hardware Corner measured)
61.4 decode tok/s (TG, single-stream, batch-1, 4K context)
Fields
id
benchmarks/comm-rtx5090-qwen3-32b-q4-decodeMetric
decode_tok_s
Value
61.4
Unit
decode tok/s (TG, single-stream, batch-1, 4K context)
Model
Qwen3-32B Q4 (GGUF)
Methodology
THIRD-PARTY MEASURED (Hardware Corner llama.cpp llama-bench). Qwen3-32B Q4: 61.4 @4K / 50.9 @16K / 43.8 @32K. Same card: Qwen3-8B Q4_K_XL TG 185.9 @4K / 169.8 @8K / 111.9 @32K, PP 10406 @4K; gpt-oss-20B 298 @4K; gpt-oss-120B MXFP4 112 @131K (entire 120B in 32GB VRAM); 35B MXFP4 97 @256K. Fastest single consumer card.
Measured by
third_party
Reproducible
yes
notes
Fastest single consumer card and simplest big-context option (32GB, 256K on MoE). New ~$3,500-3,700 (Nov-2025 $2,500 -> Feb-2026 $3,500). But $/tok-s is ~2x a used 3090 and ~equal to a 4090 - buy for one-card simplicity + max context, not as the value pick. CUDA + NVFP4 + FA2, plug and play.
Notes
Fastest single consumer card and simplest big-context option (32GB, 256K on MoE). New ~$3,500-3,700 (Nov-2025 $2,500 -> Feb-2026 $3,500). But $/tok-s is ~2x a used 3090 and ~equal to a 4090 - buy for one-card simplicity + max context, not as the value pick. CUDA + NVFP4 + FA2, plug and play.