Best local AI models for 24GB VRAM

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

A 24GB card gives you about 22.32 GiB to work with after driver overhead. 1597 indexed models fit at 32K context — the largest being Qwen3.5-99B at 99.0B parameters in I1-IQ1_S.

From the file· fit from summed bytesFrom the file· KV per layer
4GB6GB8GB10GB12GB16GB20GB24GB32GB48GB64GB96GB128GB
alltext generation 1597

Fits in 24GB at 32K context

largest quantization that fits, per model
ModelModalityBest quantParams○Total◐Headroom◐
Qwen3-Coder-30B-A3B-InstructMoEtext generationQ4_130.5B21.67 GiB0.65 GiB
Qwen3.6-27Btext generationQ5_K_M27.8B21.33 GiB0.99 GiB
Qwen3.8-27Btext generationQ5_K_M27.8B21.33 GiB0.99 GiB
gemma-4-12B-it-qat-q4_0-unquantizedtext generationQ4_012.0B9.81 GiB12.51 GiB
gemma-4-E4B-ittext generationBF168.0B15.50 GiB6.82 GiB
Qwen3-30B-A3B-Thinking-2507MoEtext generationQ4_130.5B21.69 GiB0.63 GiB
Qwen3-4Btext generationBF164.0B12.81 GiB9.51 GiB
Qwen3-8Btext generationBF168.2B20.60 GiB1.72 GiB
Laguna-XS-2.1MoEtext generationQ4_133.4B21.90 GiB0.42 GiB
Llama-3.2-1B-Instructtext generationF161.2B4.11 GiB18.21 GiB
Qwen-AgentWorld-35B-A3BMoEtext generationUD-Q4_K_M34.7B22.04 GiB0.28 GiB
gpt-oss-20bMoEtext generationF1621.5B14.40 GiB7.92 GiB
KAT-Coder-V2.5-DevMoEtext generationQ4_134.7B21.89 GiB0.43 GiB
Qwen3-30B-A3BMoEtext generationQ4_130.5B21.69 GiB0.63 GiB
llama-3-youko-8btext generationQ8_08.0B12.79 GiB9.53 GiB
Llama-3.1-8B-Instructtext generationBF168.0B19.81 GiB2.51 GiB
ced-basetext generationF3286M1.16 GiB21.16 GiB
Qwen2.5-7B-Instructtext generationF167.6B16.80 GiB5.52 GiB
UI-TARS-1.5-7Btext generationF168.3B16.80 GiB5.52 GiB
gemma-3-1b-ittext generationF161000M2.81 GiB19.51 GiB
gemma-4-E2B-it-qat-q4_0-unquantizedtext generationBF165.1B9.83 GiB12.49 GiB
Qwen3-1.7Btext generationBF162.0B8.08 GiB14.24 GiB
GLM-4.7-FlashMoEtext generationQ5_K_S31.2B21.85 GiB0.47 GiB
embeddinggemma-300mtext generationF32303M2.05 GiB20.27 GiB
Llama-3.2-3B-Instructtext generationF163.2B10.30 GiB12.02 GiB
Qwen3-0.6Btext generationBF16752M5.68 GiB16.64 GiB
Qwen3-14Btext generationQ8_014.8B20.48 GiB1.84 GiB
Ornith-1.0-35BMoEtext generationUD-Q4_K_M34.7B22.04 GiB0.28 GiB
Wan2.1-T2V-1.3Btext generationQ8_01.4B12.18 GiB10.14 GiB
Qwen3-Coder-NextMoEtext generationIQ2_XXS79.7B21.76 GiB0.56 GiB
LFM2.5-1.2B-Instructtext generationBF161.2B3.37 GiB18.95 GiB
Qwen2.5-Coder-7B-Instructtext generationQ8_07.6B17.69 GiB4.63 GiB
Qwen2.5-32B-Instructtext generationQ3_K_S32.8B22.30 GiB0.02 GiB
Qwen2.5-1.5B-Instructtext generationF161.5B4.56 GiB17.76 GiB
Jan-v3-4B-base-instructtext generationBF164.4B13.53 GiB8.79 GiB
gemma-3-4b-ittext generationBF164.3B8.85 GiB13.47 GiB
MiniCPM5-1B-Claude-Opus-Fable5-Thinkingtext generationF161.1B3.55 GiB18.77 GiB
Ornith-1.0-9Btext generationBF169.2B18.98 GiB3.34 GiB
Agents-A1MoEtext generationQ4_K_M35.1B21.14 GiB1.18 GiB
gemma-4-26B-A4B-it-ultra-uncensored-hereticMoEtext generationQ5_K_M25.8B20.15 GiB2.17 GiB
Qwen2.5-Coder-32B-Instructtext generationQ3_K_S32.8B22.30 GiB0.02 GiB
Qwen2.5-Coder-14B-Instructtext generationQ6_K_L14.8B18.49 GiB3.83 GiB
Qwen3-4B-Instruct-2507text generationF164.0B12.81 GiB9.51 GiB
granite-4.1-3btext generationBF163.4B9.65 GiB12.67 GiB
Qwen2.5-3B-Instructtext generationF323.1B13.44 GiB8.88 GiB
Phi-3.5-mini-instructtext generationQ8_03.8B16.59 GiB5.73 GiB
gemma-2-2b-ittext generationF322.6B12.41 GiB9.91 GiB
Qwen3-30B-A3B-Instruct-2507MoEtext generationQ4_130.5B21.69 GiB0.63 GiB
MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinkingtext generationF161.1B3.55 GiB18.77 GiB
Qwen2.5-0.5B-Instructtext generationF16494M2.08 GiB20.24 GiB
DeepSeek-R1-0528-Qwen3-8Btext generationBF168.2B20.60 GiB1.72 GiB
Qwen3-32Btext generationQ3_K_S32.8B22.29 GiB0.03 GiB
Qwen3-VL-8B-Instructtext generationBF168.8B20.60 GiB1.72 GiB
Sugoi-14B-Ultra-HFtext generationQ8_014.8B21.47 GiB0.85 GiB
gemma-4-12b-heretic-abliteratedtext generationQ8_012.0B15.11 GiB7.21 GiB
TinyLlama-1.1B-Chat-v1.0text generationF161.1B3.53 GiB18.79 GiB
Qwen2.5-14B-Instructtext generationQ8_014.8B21.47 GiB0.85 GiB
Mistral-Nemo-Instruct-2407text generationQ8_012.2B17.98 GiB4.34 GiB
Phi-4-mini-instructtext generationBF163.8B11.96 GiB10.36 GiB
SmolLM2-135M-Instructtext generationF16135M1.72 GiB20.60 GiB
○Spec sheet◐Predictedwhat these mean

This page models a generic 24GB 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.