Best local AI models for 32GB VRAM

Ranked by what actually fits at 32K context, computed from real file bytes.

A 32GB card gives you about 29.76 GiB to work with after driver overhead. 1673 indexed models fit at 32K context — the largest being Mixtral-8x22B-v0.1 at 141B parameters in Q6_K.

From the file· fit from summed bytesFrom the file· KV per layer
4GB6GB8GB10GB12GB16GB20GB24GB32GB48GB64GB96GB128GB
alltext generation 1673

Fits in 32GB at 32K context

largest quantization that fits, per model
ModelModalityBest quantParams○Total◐Headroom◐
Qwen3-Coder-30B-A3B-InstructMoEtext generationQ6_K30.5B27.17 GiB2.59 GiB
Qwen3.6-27Btext generationQ6_K27.8B24.18 GiB5.58 GiB
Qwen3.8-27Btext generationQ6_K27.8B24.18 GiB5.58 GiB
gemma-4-12B-it-qat-q4_0-unquantizedtext generationQ4_012.0B9.81 GiB19.95 GiB
gemma-4-E4B-ittext generationBF168.0B15.50 GiB14.26 GiB
Qwen3-30B-A3B-Thinking-2507MoEtext generationQ6_K_L30.5B27.31 GiB2.45 GiB
Qwen3-4Btext generationBF164.0B12.81 GiB16.95 GiB
Qwen3-8Btext generationBF168.2B20.60 GiB9.16 GiB
Laguna-XS-2.1MoEtext generationQ6_K_L33.4B29.30 GiB0.46 GiB
Llama-3.2-1B-Instructtext generationF161.2B4.11 GiB25.65 GiB
Qwen-AgentWorld-35B-A3BMoEtext generationUD-Q6_K34.7B28.72 GiB1.04 GiB
gpt-oss-20bMoEtext generationF1621.5B14.40 GiB15.36 GiB
KAT-Coder-V2.5-DevMoEtext generationQ6_K_L34.7B29.65 GiB0.11 GiB
Qwen3-30B-A3BMoEtext generationQ6_K_L30.5B27.31 GiB2.45 GiB
llama-3-youko-8btext generationQ8_08.0B12.79 GiB16.97 GiB
Laguna-S-2.1MoEtext generationIQ1_M118B28.21 GiB1.55 GiB
Llama-3.1-8B-Instructtext generationBF168.0B19.81 GiB9.95 GiB
ced-basetext generationF3286M1.16 GiB28.60 GiB
Qwen2.5-7B-Instructtext generationF167.6B16.80 GiB12.96 GiB
UI-TARS-1.5-7Btext generationF168.3B16.80 GiB12.96 GiB
gemma-3-1b-ittext generationF161000M2.81 GiB26.95 GiB
gemma-4-E2B-it-qat-q4_0-unquantizedtext generationBF165.1B9.83 GiB19.93 GiB
Qwen3-1.7Btext generationBF162.0B8.08 GiB21.68 GiB
GLM-4.7-FlashMoEtext generationQ6_K31.2B25.46 GiB4.30 GiB
embeddinggemma-300mtext generationF32303M2.05 GiB27.71 GiB
Llama-3.2-3B-Instructtext generationF163.2B10.30 GiB19.46 GiB
Qwen3-0.6Btext generationBF16752M5.68 GiB24.08 GiB
Qwen3-14Btext generationQ8_014.8B20.48 GiB9.28 GiB
Ornith-1.0-35BMoEtext generationQ6_K_L34.7B29.65 GiB0.11 GiB
Wan2.1-T2V-1.3Btext generationQ8_01.4B12.18 GiB17.58 GiB
Qwen3-Coder-NextMoEtext generationIQ2_M79.7B28.10 GiB1.66 GiB
LFM2.5-1.2B-Instructtext generationBF161.2B3.37 GiB26.39 GiB
Qwen2.5-Coder-7B-Instructtext generationQ8_07.6B17.69 GiB12.07 GiB
Qwen2.5-32B-Instructtext generationQ4_K_L32.8B27.93 GiB1.83 GiB
Qwen2.5-1.5B-Instructtext generationF161.5B4.56 GiB25.20 GiB
Jan-v3-4B-base-instructtext generationBF164.4B13.53 GiB16.23 GiB
gemma-3-4b-ittext generationBF164.3B8.85 GiB20.91 GiB
MiniCPM5-1B-Claude-Opus-Fable5-Thinkingtext generationF161.1B3.55 GiB26.21 GiB
Ornith-1.0-9Btext generationBF169.2B18.98 GiB10.78 GiB
Agents-A1MoEtext generationQ4_K_M35.1B21.14 GiB8.62 GiB
gemma-4-26B-A4B-it-ultra-uncensored-hereticMoEtext generationQ8_025.8B27.35 GiB2.41 GiB
Qwen2.5-Coder-32B-Instructtext generationQ4_K_L32.8B27.93 GiB1.83 GiB
Qwen2.5-Coder-14B-Instructtext generationQ6_K14.8B29.43 GiB0.33 GiB
Qwen3-4B-Instruct-2507text generationF164.0B12.81 GiB16.95 GiB
granite-4.1-3btext generationBF163.4B9.65 GiB20.11 GiB
Qwen2.5-3B-Instructtext generationF323.1B13.44 GiB16.32 GiB
Phi-3.5-mini-instructtext generationF323.8B27.04 GiB2.72 GiB
gemma-2-2b-ittext generationF322.6B12.41 GiB17.35 GiB
Qwen3-30B-A3B-Instruct-2507MoEtext generationQ6_K_L30.5B27.31 GiB2.45 GiB
MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinkingtext generationF161.1B3.55 GiB26.21 GiB
Qwen2.5-0.5B-Instructtext generationF16494M2.08 GiB27.68 GiB
DeepSeek-R1-0528-Qwen3-8Btext generationBF168.2B20.60 GiB9.16 GiB
Qwen3-32Btext generationQ4_132.8B28.11 GiB1.65 GiB
Qwen3-VL-8B-Instructtext generationBF168.8B20.60 GiB9.16 GiB
Sugoi-14B-Ultra-HFtext generationQ8_014.8B21.47 GiB8.29 GiB
gemma-4-12b-heretic-abliteratedtext generationQ8_012.0B15.11 GiB14.65 GiB
TinyLlama-1.1B-Chat-v1.0text generationF161.1B3.53 GiB26.23 GiB
Qwen2.5-14B-Instructtext generationQ8_014.8B21.47 GiB8.29 GiB
Mistral-Nemo-Instruct-2407text generationF1612.2B28.67 GiB1.09 GiB
Phi-4-mini-instructtext generationBF163.8B11.96 GiB17.80 GiB
○Spec sheet◐Predictedwhat these mean

This page models a generic 32GB accelerator, so it answers what fits rather than how fast it runs. For tokens per second you need a specific card — pick one from hardware, where bandwidth is known.