Configure a rig →
IntelDatacenterGaudi

Gaudi 2 for local AI

Gaudi 2 provides 96 GB of VRAM for local AI. In the LocalIA catalog, 227 out of 242 models run comfortably on a single card.

View all compatible models →Rig around the 2 ↗
VRAM
96GB
Category
Datacenter
Series
Gaudi
Vendor
Intel

Models that run comfortably

These models fit in 96 GB with room for context and stable inference.

★Mistral Large 123Bmistral77.3 GBcomfortableQ4 · / 96 GB
★NVIDIA Nemotron 3 Super 120B A12B BF16nemotron75.4 GBcomfortableQ4 · / 96 GB
★Llama 4 Scout 17Bx16llama68.5 GBcomfortableQ4 · / 96 GB
★Command R+ 104Bcommand79.9 GBcomfortableQ5 · / 96 GB
★Qwen3 Next 80B A3B Instructqwen61.5 GBcomfortableQ5 · / 96 GB
★Qwen 2.5 72Bqwen80.5 GBcomfortableQ8 · / 96 GB
Qwen 2.5 VL 72Bqwen80.5 GBcomfortableQ8 · / 96 GB
★Qwen2.5 72B Instructqwen80.5 GBcomfortableQ8 · / 96 GB
Llama 2 70Bllama78.2 GBcomfortableQ8 · / 96 GB
Llama 3 70Bllama78.2 GBcomfortableQ8 · / 96 GB
Llama 3.1 70Bllama78.2 GBcomfortableQ8 · / 96 GB
★Llama 3.3 70Bllama78.2 GBcomfortableQ8 · / 96 GB
CodeLlama 70Bcodellama78.2 GBcomfortableQ8 · / 96 GB
★DeepSeek R1 Distill 70Bdeepseek78.2 GBcomfortableQ8 · / 96 GB
Hermes 3 70Bhermes78.2 GBcomfortableQ8 · / 96 GB
★Llama 3.1 Nemotron 70Bnemotron78.2 GBcomfortableQ8 · / 96 GB
Athene 70Bathene78.2 GBcomfortableQ8 · / 96 GB
★Llama 3.3 70B Instructllama78.2 GBcomfortableQ8 · / 96 GB
★Llama 3.1 70B Instructllama78.2 GBcomfortableQ8 · / 96 GB
★Mixtral 8x7Bmistral52.5 GBcomfortableQ8 · / 96 GB
Falcon 40Bfalcon44.7 GBcomfortableQ8 · / 96 GB
Command R 35Bcommand78.2 GBcomfortableFP16 · / 96 GB
Aya 23 35Baya78.2 GBcomfortableFP16 · / 96 GB
CodeLlama 34Bcodellama76.0 GBcomfortableFP16 · / 96 GB
Yi 1.5 34Byi76.0 GBcomfortableFP16 · / 96 GB
★dolphin 2.9.1 yi 1.5 34byi76.0 GBcomfortableFP16 · / 96 GB
★Qwen 2.5 32Bqwen71.5 GBcomfortableFP16 · / 96 GB
★Qwen 2.5 Coder 32Bqwen71.5 GBcomfortableFP16 · / 96 GB
★Qwen 3 32Bqwen71.5 GBcomfortableFP16 · / 96 GB
★QwQ 32Bqwq71.5 GBcomfortableFP16 · / 96 GB

Tight models

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

Mixtral 8x22Bmistral88.6 GBtightQ4 · / 96 GB

Unlocked in a 2x rig

With two cards in parallel (192 GB total), larger models become reachable.

DeepSeek V2deepseek148.4 GBcomfortableQ4 · / 192 GB
DeepSeek Coder V2deepseek148.4 GBcomfortableQ4 · / 192 GB
★Qwen 3 235B A22Bqwen147.7 GBcomfortableQ4 · / 192 GB
★Qwen3 235B A22Bqwen147.7 GBcomfortableQ4 · / 192 GB
Falcon 180Bfalcon138.3 GBcomfortableQ5 · / 192 GB

Unlocked in a 4x rig

Server-style configuration (384 GB total) for the largest open-weight models.

★Llama 3.1 405Bllama311.2 GBcomfortableQ5 · / 384 GB
Hermes 3 405Bhermes311.2 GBcomfortableQ5 · / 384 GB
★Llama 4 Maverick 17Bx128llama307.3 GBcomfortableQ5 · / 384 GB
Nemotron 340Bnemotron261.2 GBcomfortableQ5 · / 384 GB

Similar GPUs

VRAM estimates updated 2026-05-12.