Can Qwen2.5-32B-Instruct run on GeForce RTX 4090 24GB (desktop)?

Conditional memory judgement for one configuration. Change context, quant, or engine and the conclusion can change.

32B Q4 on 24GB is tight and context-sensitive.

Configuration used on this page: gguf-q4_k_m, 8,192 cached tokens, 1 sequence, engine llama.cpp.

Status: Tight. Software: Documented support.

Weights 18.49 GiB · raw KV 2.00 GiB · total 20.61 GiB–21.50 GiB.

Available budget 22.00 GiB after the default reserve.

This page is an estimate unless a matching runtime-resident measurement is attached.

Full-resident weight estimates use total parameters, including inactive MoE experts.
Weight bytes come from the selected weight files only, not from every file in the repository.
Download file size is not peak GPU memory. Runtime layout, allocator padding, and KV are extra.
Counted files: qwen2.5-32b-instruct-q4_k_m-00001-of-00005.gguf, qwen2.5-32b-instruct-q4_k_m-00002-of-00005.gguf, qwen2.5-32b-instruct-q4_k_m-00003-of-00005.gguf, qwen2.5-32b-instruct-q4_k_m-00004-of-00005.gguf, qwen2.5-32b-instruct-q4_k_m-00005-of-00005.gguf.
Context budget is the number of cached tokens per sequence (prompt plus reserved generation).
Concurrent sequences share one model replica; weights are not multiplied by sequence count.
KV dtype is independent of weight quantization unless you change it.
Using 8 KV heads, not 40 query heads.

Pick one model and one device for a conditional judgement.

ModelHardwareQuantizationContext length (cached tokens)Concurrent sequences

Advanced settings

KV cache typeEngineDeploymentDevice countOverhead scenarioSystem / display reserve

GiB (2^30 bytes)

Budget already excludes system reserveExtra safety marginCustom bits per weight Include heavier prefill workspaceDisplay unit

Import local config.json

Parsed in the browser. The file is not uploaded.

Public Hugging Face repoRevision

Only huggingface.co, config.json, no token. Private repos are rejected.

Total parametersActive parametersLayersQuery headsKV headsHead dimension

Estimate

Tight

The estimate is close to the budget, or the range crosses it. Small setting changes can overflow.

Estimated memory range: 20.61 GiB – 21.50 GiB

Weights use selected file metadata. Runtime extras are still estimated.

Weights18.49 GiB

KV cache (raw)2.00 GiB

KV extras (quant metadata / packing)0.00 GiB

Activation / workspace0.06 GiB

Engine reserve0.06 GiB

Safety margin0.00 GiB

Available budget: 22.00 GiB · Remaining after estimate: 0.50 GiB – 1.39 GiB

Documented support. Enough memory does not mean this engine, quant, and OS will run. Support does not mean it will be fast.

Full-resident weight estimates use total parameters, including inactive MoE experts.
Weight bytes come from the selected weight files only, not from every file in the repository.
Download file size is not peak GPU memory. Runtime layout, allocator padding, and KV are extra.
Counted files: qwen2.5-32b-instruct-q4_k_m-00001-of-00005.gguf, qwen2.5-32b-instruct-q4_k_m-00002-of-00005.gguf, qwen2.5-32b-instruct-q4_k_m-00003-of-00005.gguf, qwen2.5-32b-instruct-q4_k_m-00004-of-00005.gguf, qwen2.5-32b-instruct-q4_k_m-00005-of-00005.gguf.
Context budget is the number of cached tokens per sequence (prompt plus reserved generation).
Concurrent sequences share one model replica; weights are not multiplied by sequence count.
KV dtype is independent of weight quantization unless you change it.
Using 8 KV heads, not 40 query heads.
Prefill peak and steady decode occupancy are not the same. The high scenario is closer to prefill/workspace pressure.
Engine reserve is not the same as bytes the model actually uses for weights and KV.
llama.cpp compute buffers grow with context and batch. Values here are user-adjustable scenarios, not measured traces — unless a matching calibration is shown.
Host RAM is not added to discrete GPU VRAM.