Best local AI models for 16GB VRAM
Ranked by what actually fits at 32K context, computed from real file bytes.
A 16GB card gives you about 14.88 GiB to work with after driver overhead. 151 indexed models fit at 32K context — the largest being Skywork-R1V3-38B at 38.4B parameters in IQ3_M.
From the file· fit from summed bytesFrom the file· KV per layer
Fits in 16GB at 32K context
largest quantization that fits, per model
| Model | Modality | Best quant | Params○ | Total◐ | Headroom◐ |
|---|---|---|---|---|---|
| Qwen3.6-35B-A3BMoE | vision + language | UD-IQ3_XXS | 36.0B | 14.53 GiB | 0.35 GiB |
| Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP | vision + language | I1-IQ3_M | 27.8B | 14.75 GiB | 0.13 GiB |
| Qwen3.5-9B | vision + language | Q8_0 | 9.7B | 10.95 GiB | 3.93 GiB |
| gemma-4-26B-A4B-itMoE | vision + language | IQ3_M | 26.5B | 14.70 GiB | 0.18 GiB |
| gemma-4-12B-it | vision + language | Q6_K_L | 12.0B | 13.08 GiB | 1.80 GiB |
| Qwen3.5-4B | vision + language | BF16 | 4.7B | 9.88 GiB | 5.00 GiB |
| Qwythos-9B-Claude-Mythos-5-1M | vision + language | Q5_K_M | 9.4B | 14.12 GiB | 0.76 GiB |
| Muse-Glimmer-30B | vision + language | UD-IQ3_M | 29.8B | 14.54 GiB | 0.34 GiB |
| Qwen3.5-0.8B | vision + language | BF16 | 873M | 2.60 GiB | 12.28 GiB |
| gemma-4-E2B-it | vision + language | BF16 | 5.1B | 9.71 GiB | 5.17 GiB |
| gemma-4-E4B-it-qat-q4_0-unquantized | vision + language | Q4_0 | 7.9B | 6.12 GiB | 8.76 GiB |
| Qwen3-VL-30B-A3B-InstructMoE | vision + language | Q2_K_L | 31.1B | 14.35 GiB | 0.53 GiB |
| Qwopus3.6-35B-A3B-v1MoE | vision + language | I1-IQ3_XXS | 36.0B | 14.12 GiB | 0.76 GiB |
| Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP | vision + language | Q4_K_M | 9.7B | 14.70 GiB | 0.18 GiB |
| Qwen3.5-35B-A3BMoE | vision + language | Q2_K_L | 36.0B | 14.47 GiB | 0.41 GiB |
| Qwythos-9B-v2 | vision + language | Q5_K_M | 9.7B | 14.16 GiB | 0.72 GiB |
| ThinkingCap-Qwen3.6-27B | vision + language | IQ3_XXS | 27.4B | 14.62 GiB | 0.26 GiB |
| Qwopus3.6-27B-Coder | vision + language | IQ2_M | 27.8B | 12.60 GiB | 2.28 GiB |
| Qwen3-VL-4B-Instruct | vision + language | BF16 | 4.4B | 12.81 GiB | 2.07 GiB |
| Qwen2.5-VL-7B-Instruct | vision + language | Q8_0 | 8.3B | 10.15 GiB | 4.73 GiB |
| Qwen3-VL-2B-Instruct | vision + language | BF16 | 2.1B | 7.50 GiB | 7.38 GiB |
| Qwen3.5-27B | vision + language | IQ3_XXS | 27.8B | 14.82 GiB | 0.06 GiB |
| gemma-3-12b-it | vision + language | Q6_K | 12.2B | 12.31 GiB | 2.57 GiB |
| Qwen3.5-2B | vision + language | BF16 | 2.3B | 4.80 GiB | 10.08 GiB |
| Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking | vision + language | IQ3_M | 27.4B | 14.87 GiB | 0.01 GiB |
| Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-DistilledMoE | vision + language | Q2_K | 36.0B | 13.77 GiB | 1.11 GiB |
| Qwen3.5-9B | vision + language | Q8_0 | 9.7B | 10.95 GiB | 3.93 GiB |
| Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved | vision + language | I1-IQ3_M | 27.4B | 14.75 GiB | 0.13 GiB |
| Mistral-Small-3.2-24B-Instruct-2506 | vision + language | Q2_K_L | 24.0B | 14.81 GiB | 0.07 GiB |
| LFM2.5-VL-1.6B | vision + language | BF16 | 1.6B | 3.37 GiB | 11.51 GiB |
| gemma-4-E4B-it-ultra-uncensored-heretic | vision + language | Q8_0 | 8.0B | 8.80 GiB | 6.08 GiB |
| Unlimited-OCRMoE | vision + language | BF16 | 3.3B | 8.14 GiB | 6.74 GiB |
| diffusiongemma-26B-A4B-itMoE | vision + language | Q3_K_M | 25.8B | 14.71 GiB | 0.17 GiB |
| Qwopus3.5-9B-v3.5 | vision + language | Q8_0 | 9.7B | 10.95 GiB | 3.93 GiB |
| Tess-4-27B | vision + language | IQ3_XXS | 27.8B | 14.62 GiB | 0.26 GiB |
| Qwen3.5-9B-GLM5.1-Distill-v1 | vision + language | Q6_K | 9.7B | 13.98 GiB | 0.90 GiB |
| Qwable-9B-Claude-Fable-5 | vision + language | Q8_0 | 9.4B | 10.71 GiB | 4.17 GiB |
| Qwen3.5-9B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING | vision + language | Q8_0 | 9.4B | 11.60 GiB | 3.28 GiB |
| Qwen3.6-35B-A3BMoE | vision + language | Q2_K | 36.0B | 13.77 GiB | 1.11 GiB |
| gemma-4-12b-it-uncensored | vision + language | I1-Q6_K | 12.0B | 12.43 GiB | 2.45 GiB |
| Holo-3.1-9B | vision + language | Q8_0 | 9.4B | 10.71 GiB | 4.17 GiB |
| MiniCPM-V-4_5 | vision + language | Q4_K_M | 8.7B | 14.70 GiB | 0.18 GiB |
| MiniCPM-V-4.6 | vision + language | BF16 | 1.3B | 2.57 GiB | 12.31 GiB |
| Qwen3.5-9B-ultra-uncensored-heretic | vision + language | Q8_0 | 9.4B | 10.71 GiB | 4.17 GiB |
| Qwen3-VL-8B-Thinking | vision + language | Q8_0 | 8.8B | 13.44 GiB | 1.44 GiB |
| Jan-v2-VL-high | vision + language | Q8_0 | 8.8B | 13.44 GiB | 1.44 GiB |
| Huihui-gemma-4-26B-A4B-it-abliteratedMoE | vision + language | UD-IQ4_NL | 26.5B | 14.83 GiB | 0.05 GiB |
| LocateAnything-3B | vision + language | F16 | 3.8B | 10.46 GiB | 4.42 GiB |
| OvisOCR2 | vision + language | F32 | 853M | 3.97 GiB | 10.91 GiB |
| gemma-3n-E2B-it | vision + language | F16 | 5.4B | 9.53 GiB | 5.35 GiB |
| Huihui-Qwen3.5-35B-A3B-abliteratedMoE | vision + language | I1-IQ3_XS | 36.0B | 14.86 GiB | 0.02 GiB |
| Qwen2-VL-2B-Instruct | vision + language | F16 | 2.2B | 4.56 GiB | 10.32 GiB |
| Qwen3-VL-30B-A3B-ThinkingMoE | vision + language | Q2_K_L | 31.1B | 14.35 GiB | 0.53 GiB |
| Gemma4-12B-Uncensored | vision + language | I1-Q6_K | 12.0B | 12.43 GiB | 2.45 GiB |
| gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic | vision + language | I1-IQ1_M | 31.3B | 14.25 GiB | 0.63 GiB |
| Fara1.5-27B | vision + language | Q2_K_L | 27.4B | 14.82 GiB | 0.06 GiB |
| Gemma4-Gutenberg-31B-Heretic | vision + language | I1-IQ1_M | 31.3B | 14.25 GiB | 0.63 GiB |
| Qwen3-VL-Embedding-2B | vision + language | F16 | 2.1B | 7.50 GiB | 7.38 GiB |
| Huihui-Qwen3-VL-4B-Instruct-abliterated | vision + language | F16 | 4.4B | 12.81 GiB | 2.07 GiB |
| pixtral-12b | vision + language | Q5_K_L | 12.7B | 14.36 GiB | 0.52 GiB |
This page models a generic 16GB 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.