Skip to content

Best LLMs for 48 GB of VRAM

What fits in 48 GB of VRAM
Text models: 184Visual models: 12Sized for RTX PRO 5000 Blackwell 48 GBStrongest first

Text models for 48 GB

Text models that fit 48 GB of VRAM, best first
ModelBest fitContextSpeedStrongest at
Qwen3.8 27B8-bit256K31 tok/sChat, Coding, Agents, Docs
Muse Glimmer 30B8-bit128K30 tok/sDocs
Qwen3.5 122B A10B2-bit32K142 tok/sChat, Coding, Agents, Docs
Nemotron 3 Super 120B A12B2-bit184K98 tok/sChat, Coding, Agents, Docs
Nemotron 3.5 Bolt 74B A9B4-bit256K89 tok/sChat, Coding, Agents, Docs
Qwen3.6 35B A3B8-bit256K171 tok/sChat, Coding, Agents, Docs
Gemma 4 31B8-bit256K27 tok/sChat, Coding, Agents, Docs
Gemma 4 26B A4B8-bit256K164 tok/sChat, Coding, Agents, Docs
GLM-4.7 Flash8-bit198K161 tok/sChat, Coding, Agents
Gemma 4 12B8-bit256K67 tok/sChat, Coding, Docs
Mistral Small 4 119B2-bit136K210 tok/sChat, Docs
gpt-oss 120B2-bit104K239 tok/sChat, Coding
Qwen3 Coder Next 80B3-bit256K221 tok/sCoding, Agents
Nemotron 3.5 Lightning 30B A3B8-bit256K152 tok/sAgents, Docs
Devstral Small 2 24BFP8128K36 tok/sCoding
LFM2.5 24B A2B8-bit128K215 tok/sAgents
Mistral Small 3.2 24B8-bit128K36 tok/sChat, Coding, Agents
Ministral 3 14B8-bit192K62 tok/sCoding, Agents, Docs
gpt-oss 20BMXFP4128K240 tok/sAgents
Ministral 3 8B8-bit256K95 tok/sDocs
Llama 3.3 70B3-bit36K26 tok/sChat, Coding, Agents, Docs
Olmo 3.1 32B8-bit64K26 tok/sChat, Coding, Agents, Docs
Llama 3.1 8B8-bit128K106 tok/sChat, Coding, Agents, Docs
Qwen3 14B8-bit40K59 tok/sNot rated
Qwen3 VL 8B Instruct8-bit256K97 tok/sText and images
Qwen3 8B8-bit40K104 tok/sNot rated
Qwen3 32B8-bit40K26 tok/sNot rated
Qwen3.6 27B8-bit256K31 tok/sText and images
Qwen2.5 Coder 14B Instruct8-bit32K58 tok/sCoding
Qwen3 Coder 30B A3B Instruct8-bit152K160 tok/sCoding
Qwen3.5 27B8-bit256K31 tok/sText and images
Qwen3.5 35B A3B8-bit256K171 tok/sText and images
NVIDIA Nemotron 3 Nano 30B A3B BF168-bit256K152 tok/sNot rated
GLM 4.5 Air2-bit32K115 tok/sNot rated
Qwen3 30B A3B8-bit40K160 tok/sNot rated
Qwen3 30B A3B Instruct 25078-bit152K160 tok/sNot rated
Laguna XS.28-bit256K172 tok/sNot rated
OTel 2.0 LLM 31B IT8-bit256K27 tok/sNot rated
Qwen2.5 32B Instruct8-bit32K26 tok/sNot rated
MiniCPM V 4.7 35B A3B8-bit256K171 tok/sText and images
Qwen2.5 14B Instruct8-bit32K58 tok/sNot rated
Laguna XS 2.18-bit256K172 tok/sNot rated
Diffusiongemma 26B A4B IT8-bit256K158 tok/sText and images
Qwen2.5 VL 32B Instruct8-bit44K26 tok/sText and images
Qwen2.5 Coder 32B Instruct8-bit32K26 tok/sCoding
Laguna S 2.12-bit72K160 tok/sNot rated
Qwen3 VL 30B A3B Instruct8-bit144K160 tok/sText and images
EXAONE 3.5 7.8B Instruct8-bit32K107 tok/sNot rated
DeepSeek R1 0528 Qwen3 8B8-bit128K104 tok/sNot rated
Gemma 3 12B IT8-bit128K71 tok/sText and images
Phi 3.5 Mini Instruct8-bit104K194 tok/sNot rated
DeepSeek R1 Distill Qwen 32B8-bit48K26 tok/sNot rated
Gemma 3 27B IT8-bit128K32 tok/sText and images
Ornith 1.5 35B A3B8-bit256K171 tok/sNot rated
Ornith 1.5 122B2-bit32K123 tok/sNot rated
GPT Neox 20B8-bit2K41 tok/sNot rated
Apertus 8B Instruct 25098-bit64K106 tok/sNot rated
Phi 48-bit16K58 tok/sNot rated
Phi 3 Mini 128K Instruct8-bit104K194 tok/sNot rated
Qwen3 VL 32B Instruct8-bit44K26 tok/sText and images

Fits, context and speed on the RTX PRO 5000 Blackwell 48 GB. A slower card with the same memory holds the same models at lower speed. Ranked by our use-case scores, then by Hugging Face downloads. The top 60 of 184 are shown.

Image and video models

Image and video models that fit 48 GB
ModelBest fitPer outputStrongest at
MiniMax H316-bit4.0 minVideo
LTX-2.516-bit42 sVideo
FLUX.1 [dev]16-bit7.9 sImages
Qwen-Image16-bit2.2 minImages
FLUX.2 [dev]8-bit60 sImages
Wan2.2 I2V A14B16-bit20 minVideo
Wan2.1 I2V 14B16-bit20 minVideo
Wan 2.1 T2V-14B16-bit25 minVideo
Krea 2 Turbo16-bit7.9 sImages
Qwen Image 251216-bit2.2 minImages
Wan 2.2 A14B16-bit20 minVideo
HunyuanVideo 1.516-bit24 minVideo

Machines with 48 GB

Questions

What is the best LLM for 48 GB of VRAM?

Of the open models we track, the strongest that run well on 48 GB are Qwen3.8 27B, Muse Glimmer 30B, Qwen3.5 122B A10B. 184 text models fit in all.

Calculated from published specs, not benchmarked. Every machine and tier