Qwen2.5-32B-Instruct: how much VRAM do you need?

The 24 GiB question is mostly about context

At 32.5 billion parameters, quantized weights take most of a 24 GiB budget before the conversation starts. This model has 64 layers, so the KV cache continues growing at a substantial rate even though the weights stay constant. A configuration that loads with a short prompt can cross the memory limit later.

Compare 4,096 and 8,192 tokens first. If you need longer contexts, inspect the Q8_0 cache option supported by your runtime and measure quality on your own prompts. Moving some weights to CPU changes the allocation, but is outside this full-GPU estimate.

A concrete 8,192-token example

With the Q4_K_M planning assumption, weights use about 18.16 GiB. One FP16 conversation cache at 8,192 tokens adds 2.00 GiB. Add the 1.5 GiB runtime allowance for an estimated 21.66 GiB allocation, before the separate display reserve. The related DeepSeek-R1-Distill-Qwen-32B preset uses 2.00 GiB of cache at those same settings. This is an architecture calculation, not a model-quality comparison.

How the estimate works

Start with the stored weights, add the conversation cache, then allow space for the runtime. The display and other applications consume a separate reserve. The result assumes the whole model and its cache stay on one GPU.

Weights (GiB) ≈ parameters × assumed bits per weight ÷ 8 ÷ 2³⁰
KV cache (GiB) = 2 × layers × KV heads × head dimension × bytes per cached value × context tokens × parallel sequences ÷ 2³⁰
Total allocation = weights + KV cache + runtime allowance

Weight presets use 4.8, 5.7, 6.6 and 8.5 effective bits per parameter for Q4_K_M, Q5_K_M, Q6_K and Q8_0. These are planning assumptions, not exact file sizes. Tensor mixtures, unquantized tensors and container metadata vary by conversion. Use the advanced weight override when you have measured the checkpoint allocation.

The cache calculation models a full dense GQA cache at the selected capacity. FP16 uses 2 bytes per value. The Q8_0 and Q4_0 estimates include block overhead (34 and 18 bytes per 32 values). Backend padding and allocation strategies can increase usage. Your runtime must support the chosen cache type and kernels; smaller caches can affect output quality.

What the fit labels mean

Usable memory is GPU capacity minus the display/other-app reserve. “Room in estimate” leaves at least 10% of that usable budget unused. “Tight budget” means the estimate fits but has less headroom. “Exceeds budget” means this full-GPU configuration exceeds the selected allocation. Partial CPU offloading may still work, with different performance.

Check it on your machine

Use the exact quantized file, runtime version and context settings you plan to deploy. Load one sequence first, inspect the runtime's memory report and GPU monitor, and then test a conversation near the intended context length. Record peak use during both prompt processing and generation before increasing concurrency.

In Ollama, ollama ps reports the processor split and allocated context. A model partially offloaded to CPU can run even when it exceeds the GPU budget. Do not interpret successful loading alone as proof of full GPU residency.

Where this tool stops

This is inference planning for the six named dense text checkpoints. It does not estimate training, fine-tuning, image generation, MLA, sliding-window caches, prefix sharing, multi-GPU placement, or Apple unified-memory limits. No tokens-per-second claims are made. Context is bounded to the presets listed in the source notes; fitting memory does not validate model quality at that length.