Best local AI models for 128GB VRAM

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

A 128GB card gives you about 119.04 GiB to work with after driver overhead. 1796 indexed models fit at 32K context — the largest being Hermes-4-405B at 406B parameters in IQ2_XXS.

From the file· fit from summed bytesFrom the file· KV per layer
4GB6GB8GB10GB12GB16GB20GB24GB32GB48GB64GB96GB128GB
alltext generation 1796

Fits in 128GB at 32K context

largest quantization that fits, per model
ModelModalityBest quantParams○Total◐Headroom◐
Qwen3-Coder-30B-A3B-InstructMoEtext generationBF1630.5B60.69 GiB58.35 GiB
Qwen3.6-27Btext generationBF1627.8B53.77 GiB65.27 GiB
Qwen3.8-27Btext generationBF1627.8B58.51 GiB60.53 GiB
DeepSeek-V4-FlashMoEtext generationUD-IQ3_S291B110.16 GiB8.88 GiB
gemma-4-12B-it-qat-q4_0-unquantizedtext generationQ4_012.0B9.81 GiB109.23 GiB
gemma-4-E4B-ittext generationBF168.0B15.50 GiB103.54 GiB
Qwen3-30B-A3B-Thinking-2507MoEtext generationBF1630.5B60.69 GiB58.35 GiB
Qwen3-4Btext generationBF164.0B12.81 GiB106.23 GiB
Qwen3-8Btext generationBF168.2B20.60 GiB98.44 GiB
Laguna-XS-2.1MoEtext generationBF1633.4B64.50 GiB54.54 GiB
DeepSeek-V4-Flash-0731MoEtext generationUD-IQ3_S304B109.01 GiB10.03 GiB
Llama-3.2-1B-Instructtext generationF161.2B4.11 GiB114.93 GiB
Qwen-AgentWorld-35B-A3BMoEtext generationBF1634.7B67.62 GiB51.42 GiB
gpt-oss-20bMoEtext generationF1621.5B14.40 GiB104.64 GiB
KAT-Coder-V2.5-DevMoEtext generationBF1634.7B66.04 GiB53.00 GiB
Qwen3-30B-A3BMoEtext generationBF1630.5B60.69 GiB58.35 GiB
llama-3-youko-8btext generationQ8_08.0B12.79 GiB106.25 GiB
Laguna-S-2.1MoEtext generationQ6_K_L118B97.30 GiB21.74 GiB
Llama-3.1-8B-Instructtext generationF328.0B34.76 GiB84.28 GiB
ced-basetext generationF3286M1.16 GiB117.88 GiB
Hy3MoEtext generationQ2_K299B112.12 GiB6.92 GiB
Qwen2.5-7B-Instructtext generationF167.6B16.80 GiB102.24 GiB
UI-TARS-1.5-7Btext generationF168.3B16.80 GiB102.24 GiB
gemma-3-1b-ittext generationF161000M2.81 GiB116.23 GiB
gemma-4-E2B-it-qat-q4_0-unquantizedtext generationBF165.1B9.83 GiB109.21 GiB
Qwen3-1.7Btext generationBF162.0B8.08 GiB110.96 GiB
GLM-4.7-FlashMoEtext generationBF1631.2B58.26 GiB60.78 GiB
embeddinggemma-300mtext generationF32303M2.05 GiB116.99 GiB
Llama-3.2-3B-Instructtext generationF163.2B10.30 GiB108.74 GiB
Qwen3-0.6Btext generationBF16752M5.68 GiB113.36 GiB
Qwen3-14Btext generationBF1614.8B33.37 GiB85.67 GiB
Ornith-1.0-35BMoEtext generationBF1634.7B67.62 GiB51.42 GiB
Wan2.1-T2V-1.3Btext generationQ8_01.4B12.18 GiB106.86 GiB
Qwen3-Coder-NextMoEtext generationQ8_079.7B82.78 GiB36.26 GiB
LFM2.5-1.2B-Instructtext generationBF161.2B3.37 GiB115.67 GiB
Qwen2.5-Coder-7B-Instructtext generationQ8_07.6B17.69 GiB101.35 GiB
Qwen2.5-32B-Instructtext generationF1632.8B69.93 GiB49.11 GiB
Qwen2.5-1.5B-Instructtext generationF161.5B4.56 GiB114.48 GiB
Jan-v3-4B-base-instructtext generationBF164.4B13.53 GiB105.51 GiB
gemma-3-4b-ittext generationBF164.3B8.85 GiB110.19 GiB
MiniCPM5-1B-Claude-Opus-Fable5-Thinkingtext generationF161.1B3.55 GiB115.49 GiB
Ornith-1.0-9Btext generationBF169.2B18.98 GiB100.06 GiB
Agents-A1MoEtext generationF1635.1B66.04 GiB53.00 GiB
gemma-4-26B-A4B-it-ultra-uncensored-hereticMoEtext generationBF1625.8B49.37 GiB69.67 GiB
Qwen2.5-Coder-32B-Instructtext generationQ8_032.8B73.76 GiB45.28 GiB
Qwen2.5-Coder-14B-Instructtext generationQ8_014.8B36.09 GiB82.95 GiB
Qwen3-4B-Instruct-2507text generationF164.0B12.81 GiB106.23 GiB
granite-4.1-3btext generationBF163.4B9.65 GiB109.39 GiB
Qwen2.5-3B-Instructtext generationF323.1B13.44 GiB105.60 GiB
Phi-3.5-mini-instructtext generationF323.8B27.04 GiB92.00 GiB
gemma-2-2b-ittext generationF322.6B12.41 GiB106.63 GiB
Qwen3-30B-A3B-Instruct-2507MoEtext generationBF1630.5B60.69 GiB58.35 GiB
MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinkingtext generationF161.1B3.55 GiB115.49 GiB
Qwen2.5-0.5B-Instructtext generationF16494M2.08 GiB116.96 GiB
DeepSeek-R1-0528-Qwen3-8Btext generationBF168.2B20.60 GiB98.44 GiB
Qwen3-32Btext generationBF1632.8B69.92 GiB49.12 GiB
Qwen3-VL-8B-Instructtext generationBF168.8B20.60 GiB98.44 GiB
Sugoi-14B-Ultra-HFtext generationF1614.8B34.36 GiB84.68 GiB
gemma-4-12b-heretic-abliteratedtext generationQ8_012.0B15.11 GiB103.93 GiB
Qwen3-235B-A22BMoEtext generationQ3_K_M235B111.43 GiB7.61 GiB
○Spec sheet◐Predictedwhat these mean

This page models a generic 128GB 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.