Which LLMs Fit in 24 GB of VRAM?

This is the 27B to 32B class. Granite Switch 4.1 30B ranks first at 22.5 GB, a tight fit. Qwen3 32B is the largest dense model that fits. At 23.9 GB it leaves almost nothing spare, so run it on a card that isn't driving your screens.

For more headroom, Muse Glimmer 30B needs 20.4 GB and Qwen 3.8 27B needs 19.2 GB. The order in the table follows the model finder's best pick, which favors newer models of a similar size. Use it to choose a few to try.

At 32K, the dense 31B and 32B models drop out. Qwen3 32B needs 31.0 GB and Gemma 4 31B (Dense) needs 24.5 GB. Muse Glimmer 30B still fits at 20.7 GB, and Qwen 3.8 27B at 21.0 GB.

The gap is in the KV cache. Qwen3 32B caches every token on all 64 of its layers, which takes 8.6 GB at 32K. Muse Glimmer 30B keeps a full cache on one layer in four and only the last 2,048 tokens on the rest, so its cache is 0.6 GB. For long documents on 24 GB, models built like that are the ones to look at.

At Q8_0, the 27B to 32B models no longer fit. Qwen 3.8 27B needs 32.9 GB there. The largest dense model at Q8_0 is Qwen3 14B, at 18.7 GB with 8K and 23.2 GB with 32K.

That makes 24 GB a Q4_K_M card for most people. A 27B to 32B model at Q4_K_M usually beats a 14B model at Q8_0. Use Q8_0 for the smaller models when you want them as close to the original as you can get.

A few models miss by less than 1 GB. Nanbeige4.3 33B needs 24.2 GB, DeepSeek R1 Distill 32B 24.2 GB and Qwen3.5 35B (MoE) 24.4 GB.

For Nanbeige4.3 33B, a q8_0 KV cache is enough. It halves the 1.8 GB cache, at a small cost in quality, and you can set it in the calculator. Dropping to Q3_K_M saves more: Nanbeige4.3 33B needs 19.8 GB there.

The RTX 3090, RTX 3090 Ti and RTX 4090 have 24 GB, and so does AMD's RX 7900 XTX. Token speed depends mostly on memory bandwidth, and the RTX 3090 and RTX 4090 are close there, so they write text at a similar pace. The RTX 4090 pulls ahead when reading long prompts, which leans on compute.

A Mac with 32 GB fits this list too. The calculator gives it 24 GB for models and keeps 8 GB for macOS and your apps, as a planning estimate. For two 24 GB cards together, see the 48 GB list.