llama3.2-3b vs qwen2.5-32b on RTX 4090

Compare llama3.2-3b and qwen2.5-32b running locally on RTX 4090 (24GB VRAM) — tokens/sec, VRAM fit, and quality scores for local inference.

Quick answer: llama3.2-3b is faster on RTX 4090 in this dataset (210 vs 44 tok/s decode).

Compare a different combination

llama3.2-3b Faster
210
tok/s (decode)
Quant Q4_K_M
VRAM required 2GB
Fits on RTX 4090
Quality score 63 / 100
Eval sources
Data Community benchmark
qwen2.5-32b
44
tok/s (decode)
Quant Q4_K_M
VRAM required 19.5GB
Fits on RTX 4090 Tight
Quality score 69 / 100
Eval sources
Data Community benchmark
Improve this GPU comparison
Measured llama3.2-3b or qwen2.5-32b on RTX 4090?

Add a source link or correction so this page can rank with real hardware-specific data instead of estimates.

Contribute GPU comparison data

GPU comparison FAQ

Which is faster on RTX 4090: llama3.2-3b or qwen2.5-32b?

llama3.2-3b is faster on RTX 4090 in this dataset (210 vs 44 tok/s decode).

Do both models fit on RTX 4090?

llama3.2-3b: ✓. qwen2.5-32b: Tight. The fit labels use Q4_K_M VRAM estimates and the benchmark data available for this GPU.

How much VRAM do these models need?

llama3.2-3b needs about 2GB in Q4_K_M; qwen2.5-32b needs about 19.5GB in Q4_K_M.

Why can speed differ by GPU?

Local LLM speed depends on GPU memory bandwidth, backend, quantization, context length, and whether the model fits entirely in VRAM or unified memory.

How can I correct this comparison?

Use the GitHub correction link on this page with your GPU, model, quantization, context length, and measured decode tokens per second.

Last updated: 2026-06-16 Suggest a correction → GitHub