Configure a rig →
AMDDatacenterInstinct CDNA 1-2

Instinct MI250X for local AI

Instinct MI250X provides 128 GB of VRAM for local AI. In the LocalIA catalog, 228 out of 242 models run comfortably on a single card.

View all compatible models →Rig around the MI250X ↗
VRAM
128GB
Category
Datacenter
Series
Instinct CDNA 1-2
Vendor
AMD

Models that run comfortably

These models fit in 128 GB with room for context and stable inference.

Mixtral 8x22Bmistral108.3 GBcomfortableQ5 · / 128 GB
★Mistral Large 123Bmistral94.5 GBcomfortableQ5 · / 128 GB
★NVIDIA Nemotron 3 Super 120B A12B BF16nemotron92.2 GBcomfortableQ5 · / 128 GB
★Llama 4 Scout 17Bx16llama83.7 GBcomfortableQ5 · / 128 GB
★Command R+ 104Bcommand79.9 GBcomfortableQ5 · / 128 GB
★Qwen3 Next 80B A3B Instructqwen89.4 GBcomfortableQ8 · / 128 GB
★Qwen 2.5 72Bqwen80.5 GBcomfortableQ8 · / 128 GB
Qwen 2.5 VL 72Bqwen80.5 GBcomfortableQ8 · / 128 GB
★Qwen2.5 72B Instructqwen80.5 GBcomfortableQ8 · / 128 GB
Llama 2 70Bllama78.2 GBcomfortableQ8 · / 128 GB
Llama 3 70Bllama78.2 GBcomfortableQ8 · / 128 GB
Llama 3.1 70Bllama78.2 GBcomfortableQ8 · / 128 GB
★Llama 3.3 70Bllama78.2 GBcomfortableQ8 · / 128 GB
CodeLlama 70Bcodellama78.2 GBcomfortableQ8 · / 128 GB
★DeepSeek R1 Distill 70Bdeepseek78.2 GBcomfortableQ8 · / 128 GB
Hermes 3 70Bhermes78.2 GBcomfortableQ8 · / 128 GB
★Llama 3.1 Nemotron 70Bnemotron78.2 GBcomfortableQ8 · / 128 GB
Athene 70Bathene78.2 GBcomfortableQ8 · / 128 GB
★Llama 3.3 70B Instructllama78.2 GBcomfortableQ8 · / 128 GB
★Llama 3.1 70B Instructllama78.2 GBcomfortableQ8 · / 128 GB
★Mixtral 8x7Bmistral105.1 GBcomfortableFP16 · / 128 GB
Falcon 40Bfalcon89.4 GBcomfortableFP16 · / 128 GB
Command R 35Bcommand78.2 GBcomfortableFP16 · / 128 GB
Aya 23 35Baya78.2 GBcomfortableFP16 · / 128 GB
CodeLlama 34Bcodellama76.0 GBcomfortableFP16 · / 128 GB
Yi 1.5 34Byi76.0 GBcomfortableFP16 · / 128 GB
★dolphin 2.9.1 yi 1.5 34byi76.0 GBcomfortableFP16 · / 128 GB
★Qwen 2.5 32Bqwen71.5 GBcomfortableFP16 · / 128 GB
★Qwen 2.5 Coder 32Bqwen71.5 GBcomfortableFP16 · / 128 GB
★Qwen 3 32Bqwen71.5 GBcomfortableFP16 · / 128 GB

Tight models

These models barely fit. They can run, but context and speed will be limited.

Falcon 180Bfalcon113.2 GBtightQ4 · / 128 GB

Unlocked in a 2x rig

With two cards in parallel (256 GB total), larger models become reachable.

★Llama 3.1 405Bllama254.6 GBtightQ4 · / 256 GB
Hermes 3 405Bhermes254.6 GBtightQ4 · / 256 GB
★Llama 4 Maverick 17Bx128llama251.5 GBtightQ4 · / 256 GB
Nemotron 340Bnemotron213.7 GBcomfortableQ4 · / 256 GB
DeepSeek V2deepseek181.3 GBcomfortableQ5 · / 256 GB
DeepSeek Coder V2deepseek181.3 GBcomfortableQ5 · / 256 GB
★Qwen 3 235B A22Bqwen180.6 GBcomfortableQ5 · / 256 GB
★Qwen3 235B A22Bqwen180.6 GBcomfortableQ5 · / 256 GB

Unlocked in a 4x rig

Server-style configuration (512 GB total) for the largest open-weight models.

★DeepSeek V3.2deepseek430.6 GBcomfortableQ4 · / 512 GB
★DeepSeek V4 Prodeepseek430.6 GBcomfortableQ4 · / 512 GB
★DeepSeek R1deepseek421.8 GBcomfortableQ4 · / 512 GB
★DeepSeek V3deepseek421.8 GBcomfortableQ4 · / 512 GB
★DeepSeek R1 (0528 snapshot)deepseek421.8 GBcomfortableQ4 · / 512 GB

Similar GPUs

VRAM estimates updated 2026-05-12.