Quick Answer: AMD Instinct MI300X offers 192GB VRAM and starts around current market pricing. It delivers approximately 1049 tokens/sec on deepseek-ai/DeepSeek-OCR. It typically draws 750W 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 AMD Instinct MI300X 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 | 1048.53 tok/sEstimated Auto-generated benchmark | 2GB |
| unsloth/Llama-3.2-3B-Instruct | Q4 | 1006.93 tok/sEstimated Auto-generated benchmark | 2GB |
| bigcode/starcoder2-3b | Q4 | 1001.74 tok/sEstimated Auto-generated benchmark | 2GB |
| tencent/HunyuanOCR | Q4 | 995.38 tok/sEstimated Auto-generated benchmark | 1GB |
| google/gemma-3-1b-it | Q4 | 965.17 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-3.2-3B-Instruct | Q4 | 964.64 tok/sEstimated Auto-generated benchmark | 2GB |
| meta-llama/Llama-3.2-1B-Instruct | Q4 | 964.27 tok/sEstimated Auto-generated benchmark | 1GB |
| context-labs/meta-llama-Llama-3.2-3B-Instruct-FP16 | Q4 | 963.56 tok/sEstimated Auto-generated benchmark | 2GB |
| google/embeddinggemma-300m | Q4 | 958.34 tok/sEstimated Auto-generated benchmark | 1GB |
| facebook/sam3 | Q4 | 954.86 tok/sEstimated Auto-generated benchmark | 1GB |
| inference-net/Schematron-3B | Q4 | 952.98 tok/sEstimated Auto-generated benchmark | 2GB |
| WeiboAI/VibeThinker-1.5B | Q4 | 951.18 tok/sEstimated Auto-generated benchmark | 1GB |
| LiquidAI/LFM2-1.2B | Q4 | 949.54 tok/sEstimated Auto-generated benchmark | 1GB |
| unsloth/Llama-3.2-1B-Instruct | Q4 | 937.51 tok/sEstimated Auto-generated benchmark | 1GB |
| google/gemma-2-2b-it | Q4 | 931.35 tok/sEstimated Auto-generated benchmark | 1GB |
| unsloth/gemma-3-1b-it | Q4 | 915.61 tok/sEstimated Auto-generated benchmark | 1GB |
| deepseek-ai/deepseek-coder-1.3b-instruct | Q4 | 913.45 tok/sEstimated Auto-generated benchmark | 2GB |
| meta-llama/Llama-3.2-1B | Q4 | 909.89 tok/sEstimated Auto-generated benchmark | 1GB |
| google-t5/t5-3b | Q4 | 906.61 tok/sEstimated Auto-generated benchmark | 2GB |
| google-bert/bert-base-uncased | Q4 | 900.66 tok/sEstimated Auto-generated benchmark | 1GB |
| TinyLlama/TinyLlama-1.1B-Chat-v1.0 | Q4 | 900.41 tok/sEstimated Auto-generated benchmark | 1GB |
| Qwen/Qwen2.5-3B-Instruct | Q4 | 888.87 tok/sEstimated Auto-generated benchmark | 2GB |
| google/gemma-2b | Q4 | 887.98 tok/sEstimated Auto-generated benchmark | 1GB |
| Qwen/Qwen2.5-3B | Q4 | 885.99 tok/sEstimated Auto-generated benchmark | 2GB |
| meta-llama/Llama-3.2-3B | Q4 | 877.19 tok/sEstimated Auto-generated benchmark | 2GB |
| meta-llama/Llama-Guard-3-1B | Q4 | 872.94 tok/sEstimated Auto-generated benchmark | 1GB |
| ibm-research/PowerMoE-3b | Q4 | 859.40 tok/sEstimated Auto-generated benchmark | 2GB |
| apple/OpenELM-1_1B-Instruct | Q4 | 853.63 tok/sEstimated Auto-generated benchmark | 1GB |
| microsoft/phi-2 | Q4 | 838.96 tok/sEstimated Auto-generated benchmark | 4GB |
| openai-community/gpt2-large | Q4 | 837.72 tok/sEstimated Auto-generated benchmark | 4GB |
| nari-labs/Dia2-2B | Q4 | 836.67 tok/sEstimated Auto-generated benchmark | 2GB |
| ibm-granite/granite-3.3-2b-instruct | Q4 | 836.40 tok/sEstimated Auto-generated benchmark | 1GB |
| deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct | Q4 | 835.86 tok/sEstimated Auto-generated benchmark | 4GB |
| allenai/OLMo-2-0425-1B | Q4 | 833.08 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Meta-Llama-3-8B | Q4 | 832.65 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen2-0.5B-Instruct | Q4 | 830.22 tok/sEstimated Auto-generated benchmark | 3GB |
| trl-internal-testing/tiny-Qwen2ForCausalLM-2.5 | Q4 | 827.44 tok/sEstimated Auto-generated benchmark | 4GB |
| black-forest-labs/FLUX.1-dev | Q4 | 827.41 tok/sEstimated Auto-generated benchmark | 4GB |
| GSAI-ML/LLaDA-8B-Instruct | Q4 | 825.90 tok/sEstimated Auto-generated benchmark | 4GB |
| deepseek-ai/DeepSeek-R1-Distill-Qwen-7B | Q4 | 825.37 tok/sEstimated Auto-generated benchmark | 4GB |
| bigscience/bloomz-560m | Q4 | 824.08 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen2-7B-Instruct | Q4 | 823.06 tok/sEstimated Auto-generated benchmark | 4GB |
| llamafactory/tiny-random-Llama-3 | Q4 | 821.96 tok/sEstimated Auto-generated benchmark | 4GB |
| meta-llama/Llama-3.1-8B | Q4 | 821.37 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen3-1.7B-Base | Q4 | 818.94 tok/sEstimated Auto-generated benchmark | 4GB |
| lmsys/vicuna-7b-v1.5 | Q4 | 815.63 tok/sEstimated Auto-generated benchmark | 4GB |
| black-forest-labs/FLUX.2-dev | Q4 | 815.57 tok/sEstimated Auto-generated benchmark | 4GB |
| microsoft/DialoGPT-small | Q4 | 815.31 tok/sEstimated Auto-generated benchmark | 4GB |
| NousResearch/Meta-Llama-3.1-8B-Instruct | Q4 | 814.40 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen2.5-0.5B-Instruct | Q4 | 811.38 tok/sEstimated Auto-generated benchmark | 3GB |
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 | Q4 | Fits comfortably | 285.07 tok/sEstimated | 17GB (have 192GB) |
| trl-internal-testing/tiny-Qwen2ForCausalLM-2.5 | FP16 | Fits comfortably | 306.61 tok/sEstimated | 15GB (have 192GB) |
| Qwen/Qwen3-4B-Instruct-2507 | Q4 | Fits comfortably | 731.87 tok/sEstimated | 2GB (have 192GB) |
| bigscience/bloomz-560m | Q8 | Fits comfortably | 539.48 tok/sEstimated | 7GB (have 192GB) |
| Qwen/Qwen3-8B | Q8 | Fits comfortably | 546.09 tok/sEstimated | 9GB (have 192GB) |
| meta-llama/Llama-2-7b-hf | Q8 | Fits comfortably | 578.76 tok/sEstimated | 7GB (have 192GB) |
| meta-llama/Llama-2-7b-hf | FP16 | Fits comfortably | 274.02 tok/sEstimated | 15GB (have 192GB) |
| deepseek-ai/DeepSeek-R1-Distill-Llama-8B | Q4 | Fits comfortably | 733.64 tok/sEstimated | 4GB (have 192GB) |
| Qwen/Qwen3-4B-Thinking-2507-FP8 | Q4 | Fits comfortably | 705.81 tok/sEstimated | 2GB (have 192GB) |
| Qwen/Qwen3-4B-Thinking-2507-FP8 | Q8 | Fits comfortably | 493.41 tok/sEstimated | 4GB (have 192GB) |
| lmsys/vicuna-7b-v1.5 | FP16 | Fits comfortably | 300.65 tok/sEstimated | 15GB (have 192GB) |
| lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-5bit | FP16 | Fits comfortably | 160.22 tok/sEstimated | 61GB (have 192GB) |
| Alibaba-NLP/gte-Qwen2-1.5B-instruct | Q4 | Fits comfortably | 778.02 tok/sEstimated | 3GB (have 192GB) |
| Qwen/Qwen2.5-32B-Instruct | Q4 | Fits comfortably | 249.75 tok/sEstimated | 17GB (have 192GB) |
| Qwen/Qwen2.5-7B-Instruct | Q8 | Fits comfortably | 516.17 tok/sEstimated | 8GB (have 192GB) |
| deepseek-ai/DeepSeek-Coder-V2-Instruct-0724 | Q4 | Fits comfortably | 114.40 tok/sEstimated | 115GB (have 192GB) |
| deepseek-ai/DeepSeek-R1-Distill-Qwen-7B | Q8 | Fits comfortably | 499.37 tok/sEstimated | 7GB (have 192GB) |
| microsoft/VibeVoice-1.5B | Q4 | Fits comfortably | 722.56 tok/sEstimated | 3GB (have 192GB) |
| openai/gpt-oss-20b | Q4 | Fits comfortably | 400.41 tok/sEstimated | 10GB (have 192GB) |
| openai/gpt-oss-20b | Q8 | Fits comfortably | 299.73 tok/sEstimated | 20GB (have 192GB) |
| openai/gpt-oss-20b | FP16 | Fits comfortably | 152.25 tok/sEstimated | 41GB (have 192GB) |
| Qwen/Qwen3-Embedding-0.6B | Q4 | Fits comfortably | 752.04 tok/sEstimated | 3GB (have 192GB) |
| Qwen/Qwen3-Embedding-0.6B | Q8 | Fits comfortably | 492.55 tok/sEstimated | 6GB (have 192GB) |
| Qwen/Qwen2.5-1.5B-Instruct | Q4 | Fits comfortably | 782.75 tok/sEstimated | 3GB (have 192GB) |
| Qwen/Qwen2.5-1.5B-Instruct | Q8 | Fits comfortably | 553.60 tok/sEstimated | 5GB (have 192GB) |
| Qwen/Qwen2.5-1.5B-Instruct | FP16 | Fits comfortably | 309.94 tok/sEstimated | 11GB (have 192GB) |
| facebook/opt-125m | Q4 | Fits comfortably | 806.00 tok/sEstimated | 4GB (have 192GB) |
| vikhyatk/moondream2 | Q4 | Fits comfortably | 703.61 tok/sEstimated | 4GB (have 192GB) |
| vikhyatk/moondream2 | Q8 | Fits comfortably | 551.02 tok/sEstimated | 7GB (have 192GB) |
| vikhyatk/moondream2 | FP16 | Fits comfortably | 313.32 tok/sEstimated | 15GB (have 192GB) |
| petals-team/StableBeluga2 | Q4 | Fits comfortably | 724.19 tok/sEstimated | 4GB (have 192GB) |
| Qwen/Qwen2.5-7B | Q4 | Fits comfortably | 731.60 tok/sEstimated | 4GB (have 192GB) |
| Qwen/Qwen2.5-7B | Q8 | Fits comfortably | 568.57 tok/sEstimated | 7GB (have 192GB) |
| Qwen/Qwen2.5-7B | FP16 | Fits comfortably | 279.34 tok/sEstimated | 15GB (have 192GB) |
| Qwen/Qwen3-1.7B | Q8 | Fits comfortably | 506.49 tok/sEstimated | 7GB (have 192GB) |
| Qwen/Qwen2.5-14B-Instruct | Q8 | Fits comfortably | 369.48 tok/sEstimated | 14GB (have 192GB) |
| Qwen/Qwen2.5-14B-Instruct | FP16 | Fits comfortably | 235.08 tok/sEstimated | 29GB (have 192GB) |
| unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit | Q4 | Fits comfortably | 774.38 tok/sEstimated | 4GB (have 192GB) |
| meta-llama/Llama-3.1-70B-Instruct | Q8 | Fits comfortably | 179.32 tok/sEstimated | 68GB (have 192GB) |
| meta-llama/Llama-3.1-70B-Instruct | FP16 | Fits comfortably | 93.00 tok/sEstimated | 137GB (have 192GB) |
| microsoft/phi-2 | Q4 | Fits comfortably | 838.96 tok/sEstimated | 4GB (have 192GB) |
| microsoft/phi-2 | Q8 | Fits comfortably | 486.37 tok/sEstimated | 7GB (have 192GB) |
| microsoft/phi-2 | FP16 | Fits comfortably | 279.01 tok/sEstimated | 15GB (have 192GB) |
| HuggingFaceTB/SmolLM2-135M | Q4 | Fits comfortably | 749.03 tok/sEstimated | 4GB (have 192GB) |
| HuggingFaceTB/SmolLM2-135M | Q8 | Fits comfortably | 501.23 tok/sEstimated | 7GB (have 192GB) |
| HuggingFaceTB/SmolLM2-135M | FP16 | Fits comfortably | 281.19 tok/sEstimated | 15GB (have 192GB) |
| zai-org/GLM-4.6-FP8 | Q4 | Fits comfortably | 765.94 tok/sEstimated | 4GB (have 192GB) |
| deepseek-ai/deepseek-coder-1.3b-instruct | FP16 | Fits comfortably | 354.75 tok/sEstimated | 6GB (have 192GB) |
| microsoft/phi-4 | Q4 | Fits comfortably | 722.09 tok/sEstimated | 4GB (have 192GB) |
| meta-llama/Llama-3.1-8B-Instruct | FP16 | Fits comfortably | 272.93 tok/sEstimated | 17GB (have 192GB) |
Note: Performance estimates are calculated. Real results may vary. Methodology · Submit real data
Explore how RTX 5070 stacks up for local inference workloads.
Explore how RTX 4060 Ti 16GB stacks up for local inference workloads.
Explore how RX 6800 XT stacks up for local inference workloads.
Explore how RTX 4070 Super stacks up for local inference workloads.
Explore how RTX 3080 stacks up for local inference workloads.