Best local AI models for 12GB VRAM
Ranked by what actually fits at 32K context, computed from real file bytes.
A 12GB card gives you about 11.16 GiB to work with after driver overhead. 1461 indexed models fit at 32K context — the largest being Skywork-R1V3-38B at 38.4B parameters in IQ2_S.
From the file· fit from summed bytesFrom the file· KV per layer
allvision + language 123speech recognition 38text generation 1238speech synthesis 21embeddings 26video generation 14image generation 1
Fits in 12GB at 32K context
largest quantization that fits, per model
| Model | Modality | Best quant | Params○ | Total◐ | Headroom◐ |
|---|---|---|---|---|---|
| Qwen3.6-35B-A3BMoE | vision + language | IQ1_M | 36.0B | 10.20 GiB | 0.96 GiB |
| Qwen3.5-9B | vision + language | Q8_0 | 9.7B | 10.95 GiB | 0.21 GiB |
| gemma-4-12B-it | vision + language | Q5_K_S | 12.0B | 11.14 GiB | 0.02 GiB |
| nemotron-3.5-asr-streaming-0.6b | speech recognition | F32 | 638M | 3.22 GiB | 7.94 GiB |
| Qwen3.5-4B | vision + language | BF16 | 4.7B | 9.88 GiB | 1.28 GiB |
| gemma-4-12B-it-qat-q4_0-unquantized | text generation | Q4_0 | 12.0B | 9.81 GiB | 1.35 GiB |
| Qwythos-9B-Claude-Mythos-5-1M | vision + language | Q5_K | 9.4B | 8.02 GiB | 3.14 GiB |
| gemma-4-E4B-it | text generation | Q8_0 | 8.0B | 8.95 GiB | 2.21 GiB |
| Qwen3-30B-A3B-Thinking-2507MoE | text generation | IQ2_XXS | 30.5B | 10.84 GiB | 0.32 GiB |
| Muse-Glimmer-30B | vision + language | IQ2_S | 29.8B | 10.73 GiB | 0.43 GiB |
| Qwen3-4B | text generation | Q8_0 | 4.0B | 9.30 GiB | 1.86 GiB |
| Qwen3-8B | text generation | Q5_K_L | 8.2B | 11.14 GiB | 0.02 GiB |
| Laguna-XS-2.1MoE | text generation | IQ2_XXS | 33.4B | 10.93 GiB | 0.23 GiB |
| Qwen3.5-0.8B | vision + language | BF16 | 873M | 2.60 GiB | 8.56 GiB |
| Llama-3.2-1B-Instruct | text generation | F16 | 1.2B | 4.11 GiB | 7.05 GiB |
| gemma-4-E2B-it | vision + language | BF16 | 5.1B | 9.71 GiB | 1.45 GiB |
| KAT-Coder-V2.5-DevMoE | text generation | IQ2_XXS | 34.7B | 10.54 GiB | 0.62 GiB |
| gemma-4-E4B-it-qat-q4_0-unquantized | vision + language | Q4_0 | 7.9B | 6.12 GiB | 5.04 GiB |
| Qwopus3.6-35B-A3B-v1MoE | vision + language | I1-IQ2_XXS | 36.0B | 10.27 GiB | 0.89 GiB |
| llama-3-youko-8b | text generation | Q5_K_M | 8.0B | 10.18 GiB | 0.98 GiB |
| Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP | vision + language | IQ2_M | 9.7B | 11.04 GiB | 0.12 GiB |
| parakeet-tdt-0.6b-v3 | speech recognition | F32 | 627M | 3.18 GiB | 7.98 GiB |
| Qwen3.5-35B-A3BMoE | vision + language | IQ1_M | 36.0B | 10.20 GiB | 0.96 GiB |
| Llama-3.1-8B-Instruct | text generation | Q6_K | 8.0B | 10.98 GiB | 0.18 GiB |
| ced-base | text generation | F32 | 86M | 1.16 GiB | 10.00 GiB |
| Ace-Step1.5 | speech synthesis | Q4_K | 160M | 9.65 GiB | 1.51 GiB |
| Qwen2.5-7B-Instruct | text generation | Q8_0 | 7.6B | 10.15 GiB | 1.01 GiB |
| Qwen3-TTS-12Hz-0.6B-Base | speech synthesis | Q8_0 | 915M | 8.68 GiB | 2.48 GiB |
| Qwythos-9B-v2 | vision + language | Q6_K_L | 9.7B | 9.59 GiB | 1.57 GiB |
| UI-TARS-1.5-7B | text generation | Q8_0 | 8.3B | 10.15 GiB | 1.01 GiB |
| gemma-3-1b-it | text generation | F16 | 1000M | 2.81 GiB | 8.35 GiB |
| gemma-4-E2B-it-qat-q4_0-unquantized | text generation | BF16 | 5.1B | 9.83 GiB | 1.33 GiB |
| Qwen3-1.7B | text generation | BF16 | 2.0B | 8.08 GiB | 3.08 GiB |
| GLM-4.7-FlashMoE | text generation | UD-IQ1_S | 31.2B | 11.07 GiB | 0.09 GiB |
| whisper-medium | speech recognition | F32 | 764M | 3.69 GiB | 7.47 GiB |
| embeddinggemma-300m | text generation | F32 | 303M | 2.05 GiB | 9.11 GiB |
| Llama-3.2-3B-Instruct | text generation | F16 | 3.2B | 10.30 GiB | 0.86 GiB |
| Qwen3-0.6B | text generation | BF16 | 752M | 5.68 GiB | 5.48 GiB |
| Qwen3-14B | text generation | UD-IQ2_M | 14.8B | 10.91 GiB | 0.25 GiB |
| Ornith-1.0-35BMoE | text generation | IQ2_XXS | 34.7B | 10.54 GiB | 0.62 GiB |
| Voxtral-Mini-4B-Realtime-2602 | speech recognition | Q8_0 | 4.4B | 8.47 GiB | 2.69 GiB |
| Wan2.1-T2V-1.3B | text generation | Q5_0 | 1.4B | 8.58 GiB | 2.58 GiB |
| LFM2.5-1.2B-Instruct | text generation | BF16 | 1.2B | 3.37 GiB | 7.79 GiB |
| Qwen2.5-Coder-7B-Instruct | text generation | Q4_0 | 7.6B | 10.86 GiB | 0.30 GiB |
| Qwen3-VL-4B-Instruct | vision + language | Q8_0 | 4.4B | 9.30 GiB | 1.86 GiB |
| Qwen2.5-1.5B-Instruct | text generation | F16 | 1.5B | 4.56 GiB | 6.60 GiB |
| Jan-v3-4B-base-instruct | text generation | Q8_0 | 4.4B | 9.68 GiB | 1.48 GiB |
| gemma-3-4b-it | text generation | BF16 | 4.3B | 8.85 GiB | 2.31 GiB |
| whisper-large-v3 | speech recognition | F16 | 1.5B | 3.74 GiB | 7.42 GiB |
| MiniCPM5-1B-Claude-Opus-Fable5-Thinking | text generation | F16 | 1.1B | 3.55 GiB | 7.61 GiB |
| Qwen2.5-VL-7B-Instruct | vision + language | Q8_0 | 8.3B | 10.15 GiB | 1.01 GiB |
| Qwen3-VL-2B-Instruct | vision + language | BF16 | 2.1B | 7.50 GiB | 3.66 GiB |
| Ornith-1.0-9B | text generation | Q8_0 | 9.2B | 10.95 GiB | 0.21 GiB |
| gemma-4-26B-A4B-it-ultra-uncensored-hereticMoE | text generation | I1-IQ2_XXS | 25.8B | 10.99 GiB | 0.17 GiB |
| jina-embeddings-v5-text-small | embeddings | F16 | 596M | 5.39 GiB | 5.77 GiB |
| whisper-large-v3-turbo | speech recognition | F16 | 809M | 2.36 GiB | 8.80 GiB |
| Qwen3-4B-Instruct-2507 | text generation | Q8_0 | 4.0B | 9.30 GiB | 1.86 GiB |
| granite-4.1-3b | text generation | BF16 | 3.4B | 9.65 GiB | 1.51 GiB |
| gemma-3-12b-it | vision + language | Q5_K_S | 12.2B | 10.98 GiB | 0.18 GiB |
| Qwen3.5-2B | vision + language | BF16 | 2.3B | 4.80 GiB | 6.36 GiB |
This page models a generic 12GB 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.