Quick Answer: Apple M3 Max offers 128GB VRAM and starts around current market pricing. It delivers approximately 70 tokens/sec on deepseek-ai/DeepSeek-OCR. It typically draws 40W under load.
This GPU offers reliable throughput for local AI workloads. Pair it with the right model quantization to hit your desired tokens/sec, and monitor prices below to catch the best deal.
Buy directly on Amazon with fast shipping and reliable customer service.
💡 Not ready to buy? Try cloud GPUs first
Test Apple M3 Max performance in the cloud before investing in hardware. Pay by the hour with no commitment.
| Model | Quantization | Tokens/sec | VRAM used |
|---|---|---|---|
| deepseek-ai/DeepSeek-OCR | Q4 | 70.03 tok/sEstimated Auto-generated benchmark | 2GB |
| google/gemma-3-1b-it | Q4 | 69.09 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-3.2-1B | Q4 | 67.81 tok/sEstimated Auto-generated benchmark | 1GB |
| google/gemma-2b | Q4 | 67.67 tok/sEstimated Auto-generated benchmark | 1GB |
| apple/OpenELM-1_1B-Instruct | Q4 | 67.33 tok/sEstimated Auto-generated benchmark | 1GB |
| ibm-granite/granite-3.3-2b-instruct | Q4 | 67.23 tok/sEstimated Auto-generated benchmark | 1GB |
| WeiboAI/VibeThinker-1.5B | Q4 | 66.35 tok/sEstimated Auto-generated benchmark | 1GB |
| Qwen/Qwen2.5-3B-Instruct | Q4 | 66.23 tok/sEstimated Auto-generated benchmark | 2GB |
| unsloth/Llama-3.2-3B-Instruct | Q4 | 65.27 tok/sEstimated Auto-generated benchmark | 2GB |
| deepseek-ai/deepseek-coder-1.3b-instruct | Q4 | 65.04 tok/sEstimated Auto-generated benchmark | 2GB |
| meta-llama/Llama-Guard-3-1B | Q4 | 64.99 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-3.2-1B-Instruct | Q4 | 64.92 tok/sEstimated Auto-generated benchmark | 1GB |
| nari-labs/Dia2-2B | Q4 | 64.71 tok/sEstimated Auto-generated benchmark | 2GB |
| context-labs/meta-llama-Llama-3.2-3B-Instruct-FP16 | Q4 | 64.62 tok/sEstimated Auto-generated benchmark | 2GB |
| LiquidAI/LFM2-1.2B | Q4 | 63.71 tok/sEstimated Auto-generated benchmark | 1GB |
| ibm-research/PowerMoE-3b | Q4 | 63.48 tok/sEstimated Auto-generated benchmark | 2GB |
| inference-net/Schematron-3B | Q4 | 63.33 tok/sEstimated Auto-generated benchmark | 2GB |
| tencent/HunyuanOCR | Q4 | 62.96 tok/sEstimated Auto-generated benchmark | 1GB |
| Qwen/Qwen2.5-3B | Q4 | 62.89 tok/sEstimated Auto-generated benchmark | 2GB |
| meta-llama/Llama-3.2-3B-Instruct | Q4 | 62.32 tok/sEstimated Auto-generated benchmark | 2GB |
| google/gemma-2-2b-it | Q4 | 61.31 tok/sEstimated Auto-generated benchmark | 1GB |
| TinyLlama/TinyLlama-1.1B-Chat-v1.0 | Q4 | 61.15 tok/sEstimated Auto-generated benchmark | 1GB |
| facebook/sam3 | Q4 | 61.06 tok/sEstimated Auto-generated benchmark | 1GB |
| unsloth/Llama-3.2-1B-Instruct | Q4 | 60.59 tok/sEstimated Auto-generated benchmark | 1GB |
| bigcode/starcoder2-3b | Q4 | 59.87 tok/sEstimated Auto-generated benchmark | 2GB |
| allenai/OLMo-2-0425-1B | Q4 | 59.36 tok/sEstimated Auto-generated benchmark | 1GB |
| HuggingFaceH4/zephyr-7b-beta | Q4 | 59.01 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen2-1.5B-Instruct | Q4 | 58.97 tok/sEstimated Auto-generated benchmark | 3GB |
| unsloth/gemma-3-1b-it | Q4 | 58.93 tok/sEstimated Auto-generated benchmark | 1GB |
| google-t5/t5-3b | Q4 | 58.85 tok/sEstimated Auto-generated benchmark | 2GB |
| Qwen/Qwen3-0.6B | Q4 | 58.82 tok/sEstimated Auto-generated benchmark | 3GB |
| lmstudio-community/Qwen3-4B-Thinking-2507-MLX-8bit | Q4 | 58.66 tok/sEstimated Auto-generated benchmark | 2GB |
| GSAI-ML/LLaDA-8B-Instruct | Q4 | 58.56 tok/sEstimated Auto-generated benchmark | 4GB |
| microsoft/Phi-3.5-vision-instruct | Q4 | 58.56 tok/sEstimated Auto-generated benchmark | 4GB |
| petals-team/StableBeluga2 | Q4 | 58.53 tok/sEstimated Auto-generated benchmark | 4GB |
| meta-llama/Llama-3.2-3B | Q4 | 58.48 tok/sEstimated Auto-generated benchmark | 2GB |
| google/embeddinggemma-300m | Q4 | 58.30 tok/sEstimated Auto-generated benchmark | 1GB |
| lmstudio-community/Qwen3-4B-Thinking-2507-MLX-4bit | Q4 | 58.12 tok/sEstimated Auto-generated benchmark | 2GB |
| meta-llama/Llama-2-7b-chat-hf | Q4 | 58.09 tok/sEstimated Auto-generated benchmark | 4GB |
| meta-llama/Meta-Llama-3-8B | Q4 | 58.06 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen2.5-1.5B | Q4 | 58.02 tok/sEstimated Auto-generated benchmark | 3GB |
| google-bert/bert-base-uncased | Q4 | 57.98 tok/sEstimated Auto-generated benchmark | 1GB |
| Qwen/Qwen2.5-0.5B-Instruct | Q4 | 57.95 tok/sEstimated Auto-generated benchmark | 3GB |
| Qwen/Qwen3-8B-FP8 | Q4 | 57.95 tok/sEstimated Auto-generated benchmark | 4GB |
| hmellor/tiny-random-LlamaForCausalLM | Q4 | 57.93 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen2.5-0.5B | Q4 | 57.88 tok/sEstimated Auto-generated benchmark | 3GB |
| Qwen/Qwen3-0.6B-Base | Q4 | 57.84 tok/sEstimated Auto-generated benchmark | 3GB |
| deepseek-ai/DeepSeek-V3-0324 | Q4 | 57.84 tok/sEstimated Auto-generated benchmark | 4GB |
| microsoft/Phi-3.5-mini-instruct | Q4 | 57.78 tok/sEstimated Auto-generated benchmark | 2GB |
| unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit | Q4 | 57.66 tok/sEstimated Auto-generated benchmark | 4GB |
Note: Performance estimates are calculated. Real results may vary. Methodology · Submit real data
| Model | Quantization | Verdict | Estimated speed | VRAM needed |
|---|---|---|---|---|
| dphn/dolphin-2.9.1-yi-1.5-34b | Q8 | Fits comfortably | 12.60 tok/sEstimated | 35GB (have 128GB) |
| Qwen/Qwen2.5-7B | FP16 | Fits comfortably | 18.80 tok/sEstimated | 15GB (have 128GB) |
| Qwen/Qwen2.5-0.5B-Instruct | Q4 | Fits comfortably | 57.95 tok/sEstimated | 3GB (have 128GB) |
| Qwen/Qwen2.5-0.5B-Instruct | FP16 | Fits comfortably | 20.40 tok/sEstimated | 11GB (have 128GB) |
| Qwen/Qwen3-30B-A3B-Instruct-2507 | Q4 | Fits comfortably | 30.16 tok/sEstimated | 15GB (have 128GB) |
| deepseek-ai/deepseek-coder-1.3b-instruct | Q8 | Fits comfortably | 43.28 tok/sEstimated | 3GB (have 128GB) |
| Qwen/Qwen3-30B-A3B | Q4 | Fits comfortably | 27.71 tok/sEstimated | 15GB (have 128GB) |
| Qwen/Qwen3-30B-A3B | Q8 | Fits comfortably | 21.92 tok/sEstimated | 31GB (have 128GB) |
| Qwen/Qwen2.5-Coder-7B-Instruct | Q8 | Fits comfortably | 38.24 tok/sEstimated | 7GB (have 128GB) |
| lmsys/vicuna-7b-v1.5 | Q4 | Fits comfortably | 53.76 tok/sEstimated | 4GB (have 128GB) |
| Qwen/Qwen2.5-3B | FP16 | Fits comfortably | 25.07 tok/sEstimated | 6GB (have 128GB) |
| nvidia/NVIDIA-Nemotron-Nano-9B-v2 | Q4 | Fits comfortably | 43.49 tok/sEstimated | 5GB (have 128GB) |
| nvidia/NVIDIA-Nemotron-Nano-9B-v2 | Q8 | Fits comfortably | 27.04 tok/sEstimated | 10GB (have 128GB) |
| dicta-il/dictalm2.0-instruct | FP16 | Fits comfortably | 21.39 tok/sEstimated | 15GB (have 128GB) |
| BSC-LT/salamandraTA-7b-instruct | Q4 | Fits comfortably | 55.16 tok/sEstimated | 4GB (have 128GB) |
| BSC-LT/salamandraTA-7b-instruct | Q8 | Fits comfortably | 35.89 tok/sEstimated | 7GB (have 128GB) |
| microsoft/Phi-3-medium-128k-instruct | FP16 | Fits comfortably | 15.52 tok/sEstimated | 29GB (have 128GB) |
| deepseek-ai/DeepSeek-OCR | Q8 | Fits comfortably | 46.07 tok/sEstimated | 4GB (have 128GB) |
| deepseek-ai/DeepSeek-V2.5 | Q4 | Not supported | 17.39 tok/sEstimated | 328GB (have 128GB) |
| deepseek-ai/DeepSeek-V2.5 | Q8 | Not supported | 14.19 tok/sEstimated | 656GB (have 128GB) |
| deepseek-ai/DeepSeek-V2.5 | FP16 | Not supported | 7.84 tok/sEstimated | 1312GB (have 128GB) |
| RedHatAI/Llama-3.2-90B-Vision-Instruct-FP8-dynamic | FP16 | Not supported | 4.30 tok/sEstimated | 176GB (have 128GB) |
| mistralai/Mistral-Large-3-675B-Instruct-2512 | Q4 | Not supported | 6.80 tok/sEstimated | 378GB (have 128GB) |
| mistralai/Mistral-Large-3-675B-Instruct-2512 | Q8 | Not supported | 4.64 tok/sEstimated | 755GB (have 128GB) |
| mistralai/Mistral-Large-3-675B-Instruct-2512 | FP16 | Not supported | 2.42 tok/sEstimated | 1509GB (have 128GB) |
| EssentialAI/rnj-1 | Q4 | Fits comfortably | 39.40 tok/sEstimated | 5GB (have 128GB) |
| EssentialAI/rnj-1 | Q8 | Fits comfortably | 29.76 tok/sEstimated | 10GB (have 128GB) |
| EssentialAI/rnj-1 | FP16 | Fits comfortably | 15.13 tok/sEstimated | 19GB (have 128GB) |
| dphn/dolphin-2.9.1-yi-1.5-34b | Q4 | Fits comfortably | 18.00 tok/sEstimated | 17GB (have 128GB) |
| dphn/dolphin-2.9.1-yi-1.5-34b | FP16 | Fits comfortably | 6.56 tok/sEstimated | 70GB (have 128GB) |
| openai/gpt-oss-20b | Q4 | Fits comfortably | 31.51 tok/sEstimated | 10GB (have 128GB) |
| openai/gpt-oss-20b | Q8 | Fits comfortably | 19.54 tok/sEstimated | 20GB (have 128GB) |
| openai/gpt-oss-20b | FP16 | Fits comfortably | 10.12 tok/sEstimated | 41GB (have 128GB) |
| google/gemma-3-1b-it | Q4 | Fits comfortably | 69.09 tok/sEstimated | 1GB (have 128GB) |
| Qwen/Qwen3-Embedding-0.6B | FP16 | Fits comfortably | 20.58 tok/sEstimated | 13GB (have 128GB) |
| Qwen/Qwen2.5-1.5B-Instruct | Q4 | Fits comfortably | 49.74 tok/sEstimated | 3GB (have 128GB) |
| facebook/opt-125m | FP16 | Fits comfortably | 19.20 tok/sEstimated | 15GB (have 128GB) |
| TinyLlama/TinyLlama-1.1B-Chat-v1.0 | Q4 | Fits comfortably | 61.15 tok/sEstimated | 1GB (have 128GB) |
| TinyLlama/TinyLlama-1.1B-Chat-v1.0 | Q8 | Fits comfortably | 41.31 tok/sEstimated | 1GB (have 128GB) |
| TinyLlama/TinyLlama-1.1B-Chat-v1.0 | FP16 | Fits comfortably | 23.39 tok/sEstimated | 2GB (have 128GB) |
| Qwen/Qwen3-4B-Instruct-2507 | Q4 | Fits comfortably | 55.54 tok/sEstimated | 2GB (have 128GB) |
| Qwen/Qwen3-8B | Q4 | Fits comfortably | 53.85 tok/sEstimated | 4GB (have 128GB) |
| Qwen/Qwen3-8B | Q8 | Fits comfortably | 34.13 tok/sEstimated | 9GB (have 128GB) |
| Qwen/Qwen3-8B | FP16 | Fits comfortably | 19.48 tok/sEstimated | 17GB (have 128GB) |
| RedHatAI/Llama-3.3-70B-Instruct-FP8-dynamic | Q4 | Fits comfortably | 18.36 tok/sEstimated | 34GB (have 128GB) |
| RedHatAI/Llama-3.3-70B-Instruct-FP8-dynamic | Q8 | Fits comfortably | 13.90 tok/sEstimated | 68GB (have 128GB) |
| meta-llama/Llama-3.1-70B-Instruct | FP16 | Not supported | 7.50 tok/sEstimated | 137GB (have 128GB) |
| microsoft/phi-2 | Q4 | Fits comfortably | 54.01 tok/sEstimated | 4GB (have 128GB) |
| microsoft/phi-2 | Q8 | Fits comfortably | 35.94 tok/sEstimated | 7GB (have 128GB) |
| openai-community/gpt2 | Q4 | Fits comfortably | 49.39 tok/sEstimated | 4GB (have 128GB) |
Note: Performance estimates are calculated. Real results may vary. Methodology · Submit real data
Explore how Apple M2 Ultra stacks up for local inference workloads.
Explore how RTX 4090 stacks up for local inference workloads.
Explore how RTX 4080 stacks up for local inference workloads.
Explore how RX 7900 XTX stacks up for local inference workloads.
Explore how NVIDIA RTX 6000 Ada stacks up for local inference workloads.
Side-by-side VRAM, throughput, efficiency, and pricing benchmarks for both GPUs.
Side-by-side VRAM, throughput, efficiency, and pricing benchmarks for both GPUs.
Side-by-side VRAM, throughput, efficiency, and pricing benchmarks for both GPUs.