Best local AI models for 64GB VRAM

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

A 64GB card gives you about 59.52 GiB to work with after driver overhead. 1761 indexed models fit at 32K context — the largest being DeepSeek-V2.5 at 236B parameters in IQ1_M.

From the file· fit from summed bytesFrom the file· KV per layer
4GB6GB8GB10GB12GB16GB20GB24GB32GB48GB64GB96GB128GB
alltext generation 1761

Fits in 64GB at 32K context

largest quantization that fits, per model
ModelModalityBest quantParams○Total◐Headroom◐
Qwen3-Coder-30B-A3B-InstructMoEtext generationQ8_030.5B34.05 GiB25.47 GiB
Qwen3.6-27Btext generationBF1627.8B53.77 GiB5.75 GiB
Qwen3.8-27Btext generationBF1627.8B58.51 GiB1.01 GiB
gemma-4-12B-it-qat-q4_0-unquantizedtext generationQ4_012.0B9.81 GiB49.71 GiB
gemma-4-E4B-ittext generationBF168.0B15.50 GiB44.02 GiB
Qwen3-30B-A3B-Thinking-2507MoEtext generationQ8_030.5B34.05 GiB25.47 GiB
Qwen3-4Btext generationBF164.0B12.81 GiB46.71 GiB
Qwen3-8Btext generationBF168.2B20.60 GiB38.92 GiB
Laguna-XS-2.1MoEtext generationQ8_033.4B35.32 GiB24.20 GiB
Llama-3.2-1B-Instructtext generationF161.2B4.11 GiB55.41 GiB
Qwen-AgentWorld-35B-A3BMoEtext generationQ8_034.7B35.80 GiB23.72 GiB
gpt-oss-20bMoEtext generationF1621.5B14.40 GiB45.12 GiB
KAT-Coder-V2.5-DevMoEtext generationQ8_034.7B35.81 GiB23.71 GiB
Qwen3-30B-A3BMoEtext generationQ8_030.5B34.05 GiB25.47 GiB
llama-3-youko-8btext generationQ8_08.0B12.79 GiB46.73 GiB
Laguna-S-2.1MoEtext generationUD-IQ4_NL118B57.18 GiB2.34 GiB
Llama-3.1-8B-Instructtext generationF328.0B34.76 GiB24.76 GiB
ced-basetext generationF3286M1.16 GiB58.36 GiB
Qwen2.5-7B-Instructtext generationF167.6B16.80 GiB42.72 GiB
UI-TARS-1.5-7Btext generationF168.3B16.80 GiB42.72 GiB
gemma-3-1b-ittext generationF161000M2.81 GiB56.71 GiB
gemma-4-E2B-it-qat-q4_0-unquantizedtext generationBF165.1B9.83 GiB49.69 GiB
Qwen3-1.7Btext generationBF162.0B8.08 GiB51.44 GiB
GLM-4.7-FlashMoEtext generationBF1631.2B58.26 GiB1.26 GiB
embeddinggemma-300mtext generationF32303M2.05 GiB57.47 GiB
Llama-3.2-3B-Instructtext generationF163.2B10.30 GiB49.22 GiB
Qwen3-0.6Btext generationBF16752M5.68 GiB53.84 GiB
Qwen3-14Btext generationBF1614.8B33.37 GiB26.15 GiB
Ornith-1.0-35BMoEtext generationQ8_034.7B35.81 GiB23.71 GiB
Wan2.1-T2V-1.3Btext generationQ8_01.4B12.18 GiB47.34 GiB
Qwen3-Coder-NextMoEtext generationUD-Q5_K_M79.7B58.96 GiB0.56 GiB
LFM2.5-1.2B-Instructtext generationBF161.2B3.37 GiB56.15 GiB
Qwen2.5-Coder-7B-Instructtext generationQ8_07.6B17.69 GiB41.83 GiB
Qwen2.5-32B-Instructtext generationQ8_032.8B41.33 GiB18.19 GiB
Qwen2.5-1.5B-Instructtext generationF161.5B4.56 GiB54.96 GiB
Jan-v3-4B-base-instructtext generationBF164.4B13.53 GiB45.99 GiB
gemma-3-4b-ittext generationBF164.3B8.85 GiB50.67 GiB
MiniCPM5-1B-Claude-Opus-Fable5-Thinkingtext generationF161.1B3.55 GiB55.97 GiB
Ornith-1.0-9Btext generationBF169.2B18.98 GiB40.54 GiB
Agents-A1MoEtext generationQ8_035.1B35.80 GiB23.72 GiB
gemma-4-26B-A4B-it-ultra-uncensored-hereticMoEtext generationBF1625.8B49.37 GiB10.15 GiB
Qwen2.5-Coder-32B-Instructtext generationQ6_K32.8B58.98 GiB0.54 GiB
Qwen2.5-Coder-14B-Instructtext generationQ8_014.8B36.09 GiB23.43 GiB
Qwen3-4B-Instruct-2507text generationF164.0B12.81 GiB46.71 GiB
granite-4.1-3btext generationBF163.4B9.65 GiB49.87 GiB
Qwen2.5-3B-Instructtext generationF323.1B13.44 GiB46.08 GiB
Phi-3.5-mini-instructtext generationF323.8B27.04 GiB32.48 GiB
gemma-2-2b-ittext generationF322.6B12.41 GiB47.11 GiB
Qwen3-30B-A3B-Instruct-2507MoEtext generationQ8_030.5B34.05 GiB25.47 GiB
MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinkingtext generationF161.1B3.55 GiB55.97 GiB
Qwen2.5-0.5B-Instructtext generationF16494M2.08 GiB57.44 GiB
DeepSeek-R1-0528-Qwen3-8Btext generationBF168.2B20.60 GiB38.92 GiB
Qwen3-32Btext generationQ8_032.8B41.32 GiB18.20 GiB
Qwen3-VL-8B-Instructtext generationBF168.8B20.60 GiB38.92 GiB
Sugoi-14B-Ultra-HFtext generationF1614.8B34.36 GiB25.16 GiB
gemma-4-12b-heretic-abliteratedtext generationQ8_012.0B15.11 GiB44.41 GiB
TinyLlama-1.1B-Chat-v1.0text generationF161.1B3.53 GiB55.99 GiB
Qwen2.5-14B-Instructtext generationF1614.8B34.36 GiB25.16 GiB
Mistral-Nemo-Instruct-2407text generationF3212.2B51.48 GiB8.04 GiB
Phi-4-mini-instructtext generationBF163.8B11.96 GiB47.56 GiB
○Spec sheet◐Predictedwhat these mean

This page models a generic 64GB 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.