Quick Answer: AMD Instinct MI250X offers 128GB VRAM and starts around current market pricing. It delivers approximately 649 tokens/sec on deepseek-ai/DeepSeek-OCR-2. It typically draws 560W 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.
With 128GB VRAM, AMD Instinct MI250X can run models up to approximately 320B parameters using 4-bit quantization. This handles most popular models including Llama 3 70B, Mistral 7B, and larger.
Consider H100 or MI300X — Maximum VRAM for enterprise workloads.
| Model | Quantization | Tokens/sec | VRAM used |
|---|---|---|---|
| deepseek-ai/DeepSeek-OCR-2 | Q4 | 649.00 tok/sEstimated Auto-generated benchmark | 2GB |
| unsloth/Llama-3.2-1B-Instruct | Q4 | 627.73 tok/sEstimated Auto-generated benchmark | 1GB |
| LiquidAI/LFM2-1.2B | Q4 | 627.55 tok/sEstimated Auto-generated benchmark | 1GB |
| google/gemma-3-1b-it | Q4 | 625.75 tok/sEstimated Auto-generated benchmark | 1GB |
| Qwen/Qwen2.5-3B | Q4 | 621.04 tok/sEstimated Auto-generated benchmark | 2GB |
| nari-labs/Dia2-2B | Q4 | 613.04 tok/sEstimated Auto-generated benchmark | 2GB |
| meta-llama/Llama-3.2-3B-Instruct | Q4 | 611.92 tok/sEstimated Auto-generated benchmark | 2GB |
| deepseek-ai/DeepSeek-OCR | Q4 | 610.41 tok/sEstimated Auto-generated benchmark | 2GB |
| google/embeddinggemma-300m | Q4 | 605.34 tok/sEstimated Auto-generated benchmark | 1GB |
| Qwen/Qwen3-ASR-1.7B | Q4 | 602.28 tok/sEstimated Auto-generated benchmark | 2GB |
| google-bert/bert-base-uncased | Q4 | 600.46 tok/sEstimated Auto-generated benchmark | 1GB |
| deepseek-ai/deepseek-coder-1.3b-instruct | Q4 | 586.19 tok/sEstimated Auto-generated benchmark | 2GB |
| nineninesix/kani-tts-2-en | Q4 | 585.96 tok/sEstimated Auto-generated benchmark | 1GB |
| facebook/sam3 | Q4 | 585.38 tok/sEstimated Auto-generated benchmark | 1GB |
| ibm-granite/granite-3.3-2b-instruct | Q4 | 582.94 tok/sEstimated Auto-generated benchmark | 1GB |
| allenai/OLMo-2-0425-1B | Q4 | 579.90 tok/sEstimated Auto-generated benchmark | 1GB |
| unsloth/Llama-3.2-3B-Instruct | Q4 | 578.12 tok/sEstimated Auto-generated benchmark | 2GB |
| bigcode/starcoder2-3b | Q4 | 570.73 tok/sEstimated Auto-generated benchmark | 2GB |
| google/gemma-2-2b-it | Q4 | 570.11 tok/sEstimated Auto-generated benchmark | 1GB |
| google/gemma-2b | Q4 | 561.32 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-Guard-3-1B | Q4 | 559.58 tok/sEstimated Auto-generated benchmark | 1GB |
| inference-net/Schematron-3B | Q4 | 558.55 tok/sEstimated Auto-generated benchmark | 2GB |
| unsloth/gemma-3-1b-it | Q4 | 557.67 tok/sEstimated Auto-generated benchmark | 1GB |
| ibm-research/PowerMoE-3b | Q4 | 551.92 tok/sEstimated Auto-generated benchmark | 2GB |
| meta-llama/Llama-3.2-1B-Instruct | Q4 | 551.64 tok/sEstimated Auto-generated benchmark | 1GB |
| Qwen/Qwen2.5-3B-Instruct | Q4 | 540.66 tok/sEstimated Auto-generated benchmark | 2GB |
| apple/OpenELM-1_1B-Instruct | Q4 | 539.42 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-3.2-1B | Q4 | 538.03 tok/sEstimated Auto-generated benchmark | 1GB |
| WeiboAI/VibeThinker-1.5B | Q4 | 527.13 tok/sEstimated Auto-generated benchmark | 1GB |
| TinyLlama/TinyLlama-1.1B-Chat-v1.0 | Q4 | 525.45 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-3.1-8B | Q4 | 524.41 tok/sEstimated Auto-generated benchmark | 4GB |
| context-labs/meta-llama-Llama-3.2-3B-Instruct-FP16 | Q4 | 524.21 tok/sEstimated Auto-generated benchmark | 2GB |
| google/gemma-3-270m-it | Q4 | 524.18 tok/sEstimated Auto-generated benchmark | 4GB |
| google-t5/t5-3b | Q4 | 524.13 tok/sEstimated Auto-generated benchmark | 2GB |
| tencent/HunyuanOCR | Q4 | 524.11 tok/sEstimated Auto-generated benchmark | 1GB |
| deepseek-ai/DeepSeek-V3.1 | Q4 | 524.05 tok/sEstimated Auto-generated benchmark | 4GB |
| trl-internal-testing/tiny-random-LlamaForCausalLM | Q4 | 523.49 tok/sEstimated Auto-generated benchmark | 4GB |
| huggyllama/llama-7b | Q4 | 522.77 tok/sEstimated Auto-generated benchmark | 4GB |
| facebook/opt-125m | Q4 | 522.55 tok/sEstimated Auto-generated benchmark | 4GB |
| microsoft/phi-4 | Q4 | 520.91 tok/sEstimated Auto-generated benchmark | 4GB |
| meta-llama/Llama-3.2-3B | Q4 | 518.65 tok/sEstimated Auto-generated benchmark | 2GB |
| Qwen/Qwen2.5-Coder-1.5B | Q4 | 517.32 tok/sEstimated Auto-generated benchmark | 3GB |
| Qwen/Qwen3-0.6B-Base | Q4 | 516.25 tok/sEstimated Auto-generated benchmark | 3GB |
| unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit | Q4 | 515.80 tok/sEstimated Auto-generated benchmark | 4GB |
| lmstudio-community/Qwen3-4B-Thinking-2507-MLX-6bit | Q4 | 515.66 tok/sEstimated Auto-generated benchmark | 2GB |
| black-forest-labs/FLUX.2-dev | Q4 | 515.42 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen2.5-7B | Q4 | 514.74 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice | Q4 | 514.72 tok/sEstimated Auto-generated benchmark | 4GB |
| ibm-granite/granite-3.3-8b-instruct | Q4 | 514.63 tok/sEstimated Auto-generated benchmark | 4GB |
| microsoft/Phi-4-multimodal-instruct | Q4 | 514.07 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 |
|---|---|---|---|---|
| openai-community/gpt2 | Q8 | Fits comfortably | 312.62 tok/sEstimated | 7GB (have 128GB) |
| openai-community/gpt2 | FP16 | Fits comfortably | 193.24 tok/sEstimated | 15GB (have 128GB) |
| Qwen/Qwen2.5-7B-Instruct | Q4 | Fits comfortably | 475.87 tok/sEstimated | 4GB (have 128GB) |
| Qwen/Qwen2.5-7B-Instruct | Q8 | Fits comfortably | 348.85 tok/sEstimated | 7GB (have 128GB) |
| Qwen/Qwen2.5-7B-Instruct | FP16 | Fits comfortably | 177.09 tok/sEstimated | 15GB (have 128GB) |
| Qwen/Qwen3-0.6B | Q4 | Fits comfortably | 476.00 tok/sEstimated | 3GB (have 128GB) |
| Qwen/Qwen3-0.6B | Q8 | Fits comfortably | 341.41 tok/sEstimated | 6GB (have 128GB) |
| Qwen/Qwen3-0.6B | FP16 | Fits comfortably | 194.29 tok/sEstimated | 13GB (have 128GB) |
| Gensyn/Qwen2.5-0.5B-Instruct | Q4 | Fits comfortably | 509.95 tok/sEstimated | 3GB (have 128GB) |
| Gensyn/Qwen2.5-0.5B-Instruct | Q8 | Fits comfortably | 334.88 tok/sEstimated | 5GB (have 128GB) |
| Gensyn/Qwen2.5-0.5B-Instruct | FP16 | Fits comfortably | 181.05 tok/sEstimated | 11GB (have 128GB) |
| meta-llama/Llama-3.1-8B-Instruct | Q4 | Fits comfortably | 489.47 tok/sEstimated | 4GB (have 128GB) |
| meta-llama/Llama-3.1-8B-Instruct | Q8 | Fits comfortably | 339.81 tok/sEstimated | 9GB (have 128GB) |
| meta-llama/Llama-3.1-8B-Instruct | FP16 | Fits comfortably | 165.15 tok/sEstimated | 17GB (have 128GB) |
| dphn/dolphin-2.9.1-yi-1.5-34b | Q4 | Fits comfortably | 160.73 tok/sEstimated | 17GB (have 128GB) |
| dphn/dolphin-2.9.1-yi-1.5-34b | Q8 | Fits comfortably | 107.15 tok/sEstimated | 35GB (have 128GB) |
| dphn/dolphin-2.9.1-yi-1.5-34b | FP16 | Fits comfortably | 65.35 tok/sEstimated | 70GB (have 128GB) |
| openai/gpt-oss-20b | Q4 | Fits comfortably | 282.04 tok/sEstimated | 10GB (have 128GB) |
| openai/gpt-oss-20b | Q8 | Fits comfortably | 181.04 tok/sEstimated | 20GB (have 128GB) |
| openai/gpt-oss-20b | FP16 | Fits comfortably | 103.74 tok/sEstimated | 41GB (have 128GB) |
| google/gemma-3-1b-it | Q4 | Fits comfortably | 625.75 tok/sEstimated | 1GB (have 128GB) |
| google/gemma-3-1b-it | Q8 | Fits comfortably | 391.47 tok/sEstimated | 1GB (have 128GB) |
| google/gemma-3-1b-it | FP16 | Fits comfortably | 218.35 tok/sEstimated | 2GB (have 128GB) |
| Qwen/Qwen3-Embedding-0.6B | Q4 | Fits comfortably | 456.09 tok/sEstimated | 3GB (have 128GB) |
| Qwen/Qwen3-Embedding-0.6B | Q8 | Fits comfortably | 331.84 tok/sEstimated | 6GB (have 128GB) |
| Qwen/Qwen3-Embedding-0.6B | FP16 | Fits comfortably | 189.24 tok/sEstimated | 13GB (have 128GB) |
| Qwen/Qwen2.5-1.5B-Instruct | Q4 | Fits comfortably | 512.33 tok/sEstimated | 3GB (have 128GB) |
| Qwen/Qwen2.5-1.5B-Instruct | Q8 | Fits comfortably | 330.17 tok/sEstimated | 5GB (have 128GB) |
| Qwen/Qwen2.5-1.5B-Instruct | FP16 | Fits comfortably | 168.04 tok/sEstimated | 11GB (have 128GB) |
| facebook/opt-125m | Q4 | Fits comfortably | 522.55 tok/sEstimated | 4GB (have 128GB) |
| facebook/opt-125m | Q8 | Fits comfortably | 324.44 tok/sEstimated | 7GB (have 128GB) |
| facebook/opt-125m | FP16 | Fits comfortably | 171.27 tok/sEstimated | 15GB (have 128GB) |
| TinyLlama/TinyLlama-1.1B-Chat-v1.0 | Q4 | Fits comfortably | 525.45 tok/sEstimated | 1GB (have 128GB) |
| TinyLlama/TinyLlama-1.1B-Chat-v1.0 | Q8 | Fits comfortably | 432.51 tok/sEstimated | 1GB (have 128GB) |
| TinyLlama/TinyLlama-1.1B-Chat-v1.0 | FP16 | Fits comfortably | 225.37 tok/sEstimated | 2GB (have 128GB) |
| trl-internal-testing/tiny-Qwen2ForCausalLM-2.5 | Q4 | Fits comfortably | 461.27 tok/sEstimated | 4GB (have 128GB) |
| trl-internal-testing/tiny-Qwen2ForCausalLM-2.5 | Q8 | Fits comfortably | 307.58 tok/sEstimated | 7GB (have 128GB) |
| trl-internal-testing/tiny-Qwen2ForCausalLM-2.5 | FP16 | Fits comfortably | 187.74 tok/sEstimated | 15GB (have 128GB) |
| Qwen/Qwen3-4B-Instruct-2507 | Q4 | Fits comfortably | 456.66 tok/sEstimated | 2GB (have 128GB) |
| Qwen/Qwen3-4B-Instruct-2507 | Q8 | Fits comfortably | 331.01 tok/sEstimated | 4GB (have 128GB) |
| Qwen/Qwen3-4B-Instruct-2507 | FP16 | Fits comfortably | 186.36 tok/sEstimated | 9GB (have 128GB) |
| meta-llama/Llama-3.2-1B-Instruct | Q4 | Fits comfortably | 551.64 tok/sEstimated | 1GB (have 128GB) |
| meta-llama/Llama-3.2-1B-Instruct | Q8 | Fits comfortably | 415.87 tok/sEstimated | 1GB (have 128GB) |
| meta-llama/Llama-3.2-1B-Instruct | FP16 | Fits comfortably | 237.03 tok/sEstimated | 2GB (have 128GB) |
| openai/gpt-oss-120b | Q4 | Fits comfortably | 97.11 tok/sEstimated | 59GB (have 128GB) |
| openai/gpt-oss-120b | Q8 | Fits comfortably | 67.62 tok/sEstimated | 117GB (have 128GB) |
| openai/gpt-oss-120b | FP16 | Not supported | 39.17 tok/sEstimated | 235GB (have 128GB) |
| Qwen/Qwen2.5-3B-Instruct | Q4 | Fits comfortably | 540.66 tok/sEstimated | 2GB (have 128GB) |
| Qwen/Qwen2.5-3B-Instruct | Q8 | Fits comfortably | 384.99 tok/sEstimated | 3GB (have 128GB) |
| openai-community/gpt2 | Q4 | Fits comfortably | 457.53 tok/sEstimated | 4GB (have 128GB) |
Note: Performance estimates are calculated. Real results may vary. Methodology · Submit real data
Buy directly on Amazon with fast shipping and reliable customer service.
Essential accessories to pair with AMD Instinct MI250X
Total Bundle Price
All items from Amazon
💡 Not ready to buy? Try cloud GPUs first
Test AMD Instinct MI250X performance in the cloud before investing in hardware. Pay by the hour with no commitment.
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.
RPG • 2020
RPG • 2023
Action RPG • 2023
RPG • 2023
Survival Horror • 2023
Action RPG • 2022
Action RPG • 2024
Action Adventure • 2025
Survival Horror • 2023
Action • 2022
Action Adventure • 2023
Action Adventure • 2019