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. 2086 indexed models fit at 32K context — the largest being MiniMax-M3 at 427B parameters in IQ2_XXS.
From the file· fit from summed bytesFrom the file· KV per layer
allvision + language 189text generation 1794speech recognition 39embeddings 26video generation 16speech synthesis 20image generation 2
Fits in 128GB at 32K context
largest quantization that fits, per model
| Model | Modality | Best quant | Params○ | Total◐ | Headroom◐ |
|---|---|---|---|---|---|
| Qwen3.6-35B-A3BMoE | vision + language | BF16 | 36.0B | 67.61 GiB | 51.43 GiB |
| embeddinggemma-300m | text generation | F32 | 303M | 2.05 GiB | 116.99 GiB |
| Qwen3.5-9B | vision + language | BF16 | 9.7B | 18.98 GiB | 100.06 GiB |
| gemma-4-26B-A4B-itMoE | vision + language | BF16 | 26.5B | 50.98 GiB | 68.06 GiB |
| nemotron-3.5-asr-streaming-0.6b | speech recognition | F32 | 638M | 3.22 GiB | 115.82 GiB |
| Qwen3-Coder-30B-A3B-InstructMoE | text generation | BF16 | 30.5B | 60.69 GiB | 58.35 GiB |
| DeepSeek-V4-FlashMoE | text generation | UD-IQ3_S | 158B | 112.79 GiB | 6.25 GiB |
| Qwen3.5-4B | vision + language | BF16 | 4.7B | 9.88 GiB | 109.16 GiB |
| Hy3MoE | text generation | Q2_K | 299B | 112.12 GiB | 6.92 GiB |
| Qwythos-9B-Claude-Mythos-5-1M | vision + language | BF16 | 9.4B | 35.67 GiB | 83.37 GiB |
| gemma-4-12B-it-qat-q4_0-unquantized | text generation | Q4_0 | 12.0B | 9.81 GiB | 109.23 GiB |
| Qwen3-VL-30B-A3B-InstructMoE | vision + language | BF16 | 31.1B | 60.69 GiB | 58.35 GiB |
| gemma-4-26B-A4B-it-qat-q4_0-unquantizedMoE | vision + language | Q4_0 | 26.5B | 15.78 GiB | 103.26 GiB |
| Llama-3.2-1B-Instruct | text generation | F16 | 1.2B | 4.11 GiB | 114.93 GiB |
| Qwen-AgentWorld-35B-A3BMoE | text generation | BF16 | 34.7B | 67.62 GiB | 51.42 GiB |
| llama-3-youko-8b | text generation | Q8_0 | 8.0B | 12.79 GiB | 106.25 GiB |
| Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP | vision + language | Q6_K | 27.8B | 90.87 GiB | 28.17 GiB |
| gpt-oss-20bMoE | text generation | F16 | 21.5B | 14.40 GiB | 104.64 GiB |
| Qwen3-8B | text generation | BF16 | 8.2B | 20.60 GiB | 98.44 GiB |
| Qwopus3.6-35B-A3B-v1MoE | vision + language | F16 | 36.0B | 67.61 GiB | 51.43 GiB |
| Qwen3.5-0.8B | vision + language | BF16 | 873M | 2.60 GiB | 116.44 GiB |
| gemma-4-e4b-it | vision + language | BF16 | 8.0B | 15.50 GiB | 103.54 GiB |
| gemma-4-31B-it-qat-q4_0-unquantized | vision + language | Q4_0 | 32.7B | 23.49 GiB | 95.55 GiB |
| Qwen3-4B | text generation | BF16 | 4.0B | 12.81 GiB | 106.23 GiB |
| UI-TARS-1.5-7B | text generation | F16 | 8.3B | 16.80 GiB | 102.24 GiB |
| Llama-3.1-8B-Instruct | text generation | F32 | 8.0B | 34.76 GiB | 84.28 GiB |
| Ornith-1.0-35BMoE | text generation | BF16 | 34.7B | 67.62 GiB | 51.42 GiB |
| parakeet-tdt-0.6b-v3 | speech recognition | F32 | 627M | 3.18 GiB | 115.86 GiB |
| ThinkingCap-Qwen3.6-27B | vision + language | F16 | 27.4B | 53.77 GiB | 65.27 GiB |
| Llama-3.2-3B-Instruct | text generation | F16 | 3.2B | 10.30 GiB | 108.74 GiB |
| Qwopus3.6-27B-Coder | vision + language | Q8_0 | 27.8B | 29.92 GiB | 89.12 GiB |
| whisper-medium | speech recognition | F32 | 764M | 3.69 GiB | 115.35 GiB |
| Qwythos-9B-v2 | vision + language | BF16 | 9.7B | 35.67 GiB | 83.37 GiB |
| gemma-4-E4B-it-qat-q4_0-unquantized | vision + language | Q4_0 | 7.9B | 6.12 GiB | 112.92 GiB |
| Voxtral-Mini-4B-Realtime-2602 | speech recognition | F16 | 4.4B | 12.33 GiB | 106.71 GiB |
| Qwen3.5-122B-A10BMoE | vision + language | Q6_K_L | 125B | 102.88 GiB | 16.16 GiB |
| Wan2.1-T2V-1.3B | text generation | Q8_0 | 1.4B | 12.18 GiB | 106.86 GiB |
| Qwen3-Coder-NextMoE | text generation | Q8_0 | 79.7B | 82.78 GiB | 36.26 GiB |
| LFM2.5-1.2B-Instruct | text generation | BF16 | 1.2B | 3.37 GiB | 115.67 GiB |
| gemma-3-1b-it | text generation | F16 | 1000M | 2.81 GiB | 116.23 GiB |
| Qwen2.5-7B-Instruct | text generation | F16 | 7.6B | 16.80 GiB | 102.24 GiB |
| Qwen3-14B | text generation | BF16 | 14.8B | 33.37 GiB | 85.67 GiB |
| Qwen3-0.6B | text generation | BF16 | 752M | 5.68 GiB | 113.36 GiB |
| Qwen3.5-35B-A3BMoE | vision + language | BF16 | 36.0B | 67.61 GiB | 51.43 GiB |
| Qwen2.5-Coder-7B-Instruct | text generation | Q8_0 | 7.6B | 17.69 GiB | 101.35 GiB |
| Qwen2.5-32B-Instruct | text generation | F16 | 32.8B | 69.93 GiB | 49.11 GiB |
| Qwen3-VL-4B-Instruct | vision + language | BF16 | 4.4B | 12.81 GiB | 106.23 GiB |
| Qwen2.5-1.5B-Instruct | text generation | F16 | 1.5B | 4.56 GiB | 114.48 GiB |
| Jan-v3-4B-base-instruct | text generation | BF16 | 4.4B | 13.53 GiB | 105.51 GiB |
| gemma-3-4b-it | text generation | BF16 | 4.3B | 8.85 GiB | 110.19 GiB |
| Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking | vision + language | Q8_0 | 39.5B | 43.77 GiB | 75.27 GiB |
| whisper-large-v3 | speech recognition | F16 | 1.5B | 3.74 GiB | 115.30 GiB |
| MiniCPM5-1B-Claude-Opus-Fable5-Thinking | text generation | F16 | 1.1B | 3.55 GiB | 115.49 GiB |
| Qwen3-30B-A3BMoE | text generation | BF16 | 30.5B | 60.69 GiB | 58.35 GiB |
| Qwen2.5-VL-7B-Instruct | vision + language | BF16 | 8.3B | 16.80 GiB | 102.24 GiB |
| Qwen3-VL-2B-Instruct | vision + language | BF16 | 2.1B | 7.50 GiB | 111.54 GiB |
| Ornith-1.0-9B | text generation | BF16 | 9.2B | 18.98 GiB | 100.06 GiB |
| Qwen3-1.7B | text generation | BF16 | 2.0B | 7.50 GiB | 111.54 GiB |
| Laguna-S-2.1MoE | text generation | Q6_K_L | 118B | 97.30 GiB | 21.74 GiB |
| Agents-A1MoE | text generation | F16 | 35.1B | 66.04 GiB | 53.00 GiB |
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.