Configure a rig →
NVIDIAWorkstationRTX Ada

RTX 4000 Ada for local AI

RTX 4000 Ada provides 20 GB of VRAM for local AI. In the LocalIA catalog, 179 out of 242 models run comfortably on a single card.

View all compatible models →Rig around the Ada ↗
VRAM
20GB
Category
Workstation
Series
RTX Ada
Vendor
NVIDIA

Models that run comfortably

These models fit in 20 GB with room for context and stable inference.

★Gemma 2 27Bgemma17.0 GBcomfortableQ4 · / 20 GB
★Gemma 3 27Bgemma17.0 GBcomfortableQ4 · / 20 GB
★Gemma 4 26B A4Bgemma16.3 GBcomfortableQ4 · / 20 GB
★Mistral Small 3 24Bmistral15.1 GBcomfortableQ4 · / 20 GB
★Mistral Small 3.1 24Bmistral15.1 GBcomfortableQ4 · / 20 GB
★Mistral Small 3.2 24Bmistral15.1 GBcomfortableQ4 · / 20 GB
★Devstral Small 2 24Bdevstral15.1 GBcomfortableQ4 · / 20 GB
Mistral Small 22Bmistral16.9 GBcomfortableQ5 · / 20 GB
★Codestral 22Bcodestral16.9 GBcomfortableQ5 · / 20 GB
Reka Flash 3 21Breka16.1 GBcomfortableQ5 · / 20 GB
InternLM 2.5 20Binternlm15.4 GBcomfortableQ5 · / 20 GB
DeepSeek V2 Litedeepseek12.3 GBcomfortableQ5 · / 20 GB
DeepSeek Coder V2 Litedeepseek12.3 GBcomfortableQ5 · / 20 GB
StarCoder 2 15Bstarcoder16.8 GBcomfortableQ8 · / 20 GB
★Phi-4 Reasoning Vision 15Bphi16.8 GBcomfortableQ8 · / 20 GB
★Qwen 2.5 14Bqwen15.6 GBcomfortableQ8 · / 20 GB
Qwen 2.5 Coder 14Bqwen15.6 GBcomfortableQ8 · / 20 GB
★Qwen 3 14Bqwen15.6 GBcomfortableQ8 · / 20 GB
★DeepSeek R1 Distill 14Bdeepseek15.6 GBcomfortableQ8 · / 20 GB
Phi-3 Medium 14Bphi15.6 GBcomfortableQ8 · / 20 GB
★Phi-4 14Bphi15.6 GBcomfortableQ8 · / 20 GB
★GLM-4.5 Airglm15.6 GBcomfortableQ8 · / 20 GB
★Qwen2.5 14B Instructqwen15.6 GBcomfortableQ8 · / 20 GB
★Qwen3 14Bqwen15.6 GBcomfortableQ8 · / 20 GB
★Qwen2.5 Coder 14B Instructqwen15.6 GBcomfortableQ8 · / 20 GB
★DeepSeek R1 Distill Qwen 14Bqwen15.6 GBcomfortableQ8 · / 20 GB
Llama 2 13Bllama14.5 GBcomfortableQ8 · / 20 GB
CodeLlama 13Bcodellama14.5 GBcomfortableQ8 · / 20 GB
OLMo 2 13Bolmo14.5 GBcomfortableQ8 · / 20 GB
Vicuna 13Bvicuna14.5 GBcomfortableQ8 · / 20 GB

Tight models

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

★Gemma 4 31Bgemma19.5 GBtightQ4 · / 20 GB
★Qwen 3 30B A3Bqwen18.9 GBtightQ4 · / 20 GB
MPT 30Bmpt18.9 GBtightQ4 · / 20 GB
★Qwen3 Coder 30B A3B Instructqwen18.9 GBtightQ4 · / 20 GB
★Qwen3 30B A3Bqwen18.9 GBtightQ4 · / 20 GB
★Qwen3 30B A3B Instruct 2507qwen18.9 GBtightQ4 · / 20 GB
★NVIDIA Nemotron 3 Nano 30B A3B BF16nemotron18.9 GBtightQ4 · / 20 GB

Unlocked in a 2x rig

With two cards in parallel (40 GB total), larger models become reachable.

★Mixtral 8x7Bmistral29.5 GBcomfortableQ4 · / 40 GB
Falcon 40Bfalcon30.7 GBcomfortableQ5 · / 40 GB
Command R 35Bcommand26.9 GBcomfortableQ5 · / 40 GB
Aya 23 35Baya26.9 GBcomfortableQ5 · / 40 GB
CodeLlama 34Bcodellama26.1 GBcomfortableQ5 · / 40 GB
Yi 1.5 34Byi26.1 GBcomfortableQ5 · / 40 GB
★dolphin 2.9.1 yi 1.5 34byi26.1 GBcomfortableQ5 · / 40 GB
★Qwen 2.5 32Bqwen24.6 GBcomfortableQ5 · / 40 GB
★Qwen 2.5 Coder 32Bqwen24.6 GBcomfortableQ5 · / 40 GB
★Qwen 3 32Bqwen24.6 GBcomfortableQ5 · / 40 GB
★QwQ 32Bqwq24.6 GBcomfortableQ5 · / 40 GB
★DeepSeek R1 Distill 32Bdeepseek24.6 GBcomfortableQ5 · / 40 GB
Qwen 2.5 VL 32Bqwen24.6 GBcomfortableQ5 · / 40 GB
★Granite 4 H-Small 32B-A9Bgranite24.6 GBcomfortableQ5 · / 40 GB
GLM-4.6glm24.6 GBcomfortableQ5 · / 40 GB

Unlocked in a 4x rig

Server-style configuration (80 GB total) for the largest open-weight models.

★Mistral Large 123Bmistral77.3 GBtightQ4 · / 80 GB
★NVIDIA Nemotron 3 Super 120B A12B BF16nemotron75.4 GBtightQ4 · / 80 GB
★Llama 4 Scout 17Bx16llama68.5 GBtightQ4 · / 80 GB
★Command R+ 104Bcommand65.4 GBcomfortableQ4 · / 80 GB
★Qwen3 Next 80B A3B Instructqwen61.5 GBcomfortableQ5 · / 80 GB
★Qwen 2.5 72Bqwen55.3 GBcomfortableQ5 · / 80 GB
Qwen 2.5 VL 72Bqwen55.3 GBcomfortableQ5 · / 80 GB
★Qwen2.5 72B Instructqwen55.3 GBcomfortableQ5 · / 80 GB
Llama 2 70Bllama53.8 GBcomfortableQ5 · / 80 GB
Llama 3 70Bllama53.8 GBcomfortableQ5 · / 80 GB

Similar GPUs

VRAM estimates updated 2026-05-12.