RigPerDollar

16GB vs 24GB vs 32GB VRAM for local AI

16 GB covers 31 of our 79 current models. The cheapest 16 GB card is the Radeon RX 9070 at $570; 24 GB (from $1,350) fits 51 and 32 GB (from $6,500) fits 53.

What each memory size runs from our list of current models, what the next step up adds, and what the cheapest card at each size costs today.

VRAMCurrent models that fitLargest that fitsCheapest priced cardIts speed on that model estimate$ per GB of VRAM
16 GB31 of 79Mistral-Small-3.2-24B-Instruct-2506 (14.33 GB)Radeon RX 9070 $570 (new)~34 tok/s$36
24 GB51 of 79K2-Horizon-MoVA-36B-A4B (22.37 GB, MoE)Radeon RX 7900 XTX $1,350 (new)~178 tok/s$56
32 GB53 of 79Recursive Flint 40B-A2.5B (25.99 GB (Q4_1), MoE)GeForce RTX 5090 $6,500 (new)~366 tok/s$203

16 GB: fits 31 of our 79 current models; the step up from 12 GB adds 3 — Mistral-Small-3.2-24B-Instruct-2506, Devstral-Small-2505, gpt-oss-20b. Cheapest card: AMD Radeon RX 9070, $570.

16 → 24 GB: adds 20 models — K2-Horizon-MoVA-36B-A4B, Qwen3.6-35B-A3B, Qwen3.5-35B-A3B, NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB (2-bit) and 16 more. Cheapest card: AMD Radeon RX 7900 XTX, $1,350 ($780 more than the cheapest 16 GB card).

24 → 32 GB: adds 2 models — Recursive Flint 40B-A2.5B, NVIDIA Nemotron 3.5 Lightning 30B-A3B. Cheapest card: NVIDIA GeForce RTX 5090, $6,500 ($5,150 more than the cheapest 24 GB card).

Beyond 32 GB: 26 of our 79 current models need more than any single card here, from Ripple-65B Preview up. Those are multi-GPU or unified-memory territory.

Last updated Oct 16, 2026. Speeds are bandwidth-based estimates, not benchmarks. Prices are dated snapshots and move daily; "No price yet" means we have no verified price.

Which current models fit at each size

ModelFileNeeds16 GB24 GB32 GBOn the cheapest card that fits estimate
gpt-oss-20b MoE11.67 GB13.17 GBFitsFitsFits~154 tok/s on Radeon RX 9070
Devstral-Small-250514.33 GB15.83 GBTightFitsFits~34 tok/s on Radeon RX 9070
Mistral-Small-3.2-24B-Instruct-250614.33 GB15.83 GBTightFitsFits~34 tok/s on Radeon RX 9070
LFM2.5-24B-A2B MoE14.53 GB16.03 GBNoFitsFits~155 tok/s on Radeon RX 7900 XTX
Kvark Sparkle 1.1 27B16.81 GB18.31 GBNoFitsFits~40 tok/s on Radeon RX 7900 XTX
Gemma 4 26B A4B MoE17.04 GB18.54 GBNoFitsFits~173 tok/s on Radeon RX 7900 XTX
Qwen3.8-27B17.44 GB18.94 GBNoFitsFits~38 tok/s on Radeon RX 7900 XTX
Microsoft Fara1.5 27B17.53 GB19.03 GBNoFitsFits~38 tok/s on Radeon RX 7900 XTX
IBM Granite 4.2 30B17.72 GB19.22 GBNoFitsFits~42 tok/s on Radeon RX 7900 XTX
Qwen3.5-27B17.98 GB19.48 GBNoFitsFits~37 tok/s on Radeon RX 7900 XTX
Qwen3.6-27B17.98 GB19.48 GBNoFitsFits~37 tok/s on Radeon RX 7900 XTX
GLM-4.7-Flash MoE18.31 GB19.81 GBNoFitsFits~209 tok/s on Radeon RX 7900 XTX
Gemma 4 31B18.32 GB19.82 GBNoFitsFits~40 tok/s on Radeon RX 7900 XTX
Meta Muse Glimmer 30B18.39 GB19.89 GBNoFitsFits~40 tok/s on Radeon RX 7900 XTX
Raon-1 mini preview MoE19.69 GB21.19 GBNoFitsFits~114 tok/s on Radeon RX 7900 XTX
Xing4.0-29B-A4B MoE20.06 GB21.56 GBNoFitsFits~170 tok/s on Radeon RX 7900 XTX
Nanbeige4.3-33B20.79 GB22.29 GBNoFitsFits~36 tok/s on Radeon RX 7900 XTX
K2-Horizon-32B21.08 GB22.58 GBNoTightFits~35 tok/s on Radeon RX 7900 XTX
MiniCPM-V 4.7 35B-A3B MoE21.17 GB22.67 GBNoFitsFits~138 tok/s on Radeon RX 7900 XTX
NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB (2-bit) MoE22.02 GB (native 2-bit)23.52 GBNoFitsFits~119 tok/s on Radeon RX 7900 XTX
Qwen3.5-35B-A3B MoE22.29 GB23.79 GBNoFitsFits~137 tok/s on Radeon RX 7900 XTX
Qwen3.6-35B-A3B MoE22.29 GB23.79 GBNoFitsFits~137 tok/s on Radeon RX 7900 XTX
K2-Horizon-MoVA-36B-A4B MoE22.37 GB23.87 GBNoTightFits~178 tok/s on Radeon RX 7900 XTX
NVIDIA Nemotron 3.5 Lightning 30B-A3B MoE25.48 GB26.98 GBNoNoFits~177 tok/s on GeForce RTX 5090
Recursive Flint 40B-A2.5B MoE25.99 GB (Q4_1)27.49 GBNoNoFits~366 tok/s on GeForce RTX 5090

Current models that don't fit 12 GB but do fit 32 GB, smallest first. Fit is the file plus 1.5 GB of headroom against the card's total VRAM and at least 4K tokens of context fit with an 8-bit KV cache (details); "Tight" means room for under 8K tokens of context with the default FP16 KV cache, so short contexts only. MoE models read only their active experts per token, which is why their estimates are so much higher. Unified-memory Macs and the Framework Desktop are left out because their memory isn't one of these sizes; see Mac vs PC.

Who this is for

You're choosing how much VRAM to pay for. Pick the smallest size that fits the models you'll actually run, then shop on price per token per second within that size: the GPU ranking sorts every card, and the 70B guide covers what lies past 32 GB.

Questions people ask

How much VRAM do you need: 16 GB, 24 GB or 32 GB?

16 GB covers 31 of our 79 current models. The cheapest 16 GB card is the Radeon RX 9070 at $570; 24 GB (from $1,350) fits 51 and 32 GB (from $6,500) fits 53.

What does 24 GB of VRAM run that 16 GB can't?

20 of our 79 current models: K2-Horizon-MoVA-36B-A4B (22.37 GB), Qwen3.6-35B-A3B (22.29 GB), Qwen3.5-35B-A3B (22.29 GB) and NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB (2-bit) (22.02 GB), and 16 more. The cheapest priced 24 GB card is the AMD Radeon RX 7900 XTX at $1,350 (new), against $570 for the cheapest 16 GB card.

What does 32 GB of VRAM run that 24 GB can't?

2 of our 79 current models: Recursive Flint 40B-A2.5B (25.99 GB) and NVIDIA Nemotron 3.5 Lightning 30B-A3B (25.48 GB). The cheapest priced 32 GB card is the NVIDIA GeForce RTX 5090 at $6,500 (new), against $1,350 for the cheapest 24 GB card.

Is 16 GB of VRAM enough for local LLMs?

For 31 of our 79 current models, yes, up to Mistral-Small-3.2-24B-Instruct-2506 (14.33 GB at Q4_K_M). K2-Horizon-MoVA-36B-A4B, Qwen3.6-35B-A3B and Qwen3.5-35B-A3B need 24 GB, and 26 current models need more than 32 GB.

Related