Skip to content

Best LLMs for 16 GB of VRAM

What fits in 16 GB of VRAM
Text models: 247Visual models: 18Sized for RTX 5060 Ti 16 GBStrongest first

Text models for 16 GB

Text models that fit 16 GB of VRAM, best first
ModelBest fitContextSpeedStrongest at
Qwen3.8 27B3-bit20K23 tok/sChat, Coding, Agents, Docs
Muse Glimmer 30B3-bit40K22 tok/sDocs
Gemma 4 26B A4B3-bit224K129 tok/sChat, Coding, Agents, Docs
Qwen3.5 9B8-bit144K33 tok/sChat, Coding, Agents, Docs
Gemma 4 12B8-bit248K24 tok/sChat, Coding, Docs
Qwen3.5 4B8-bit256K59 tok/sAgents, Docs
Devstral Small 2 24B3-bit20K27 tok/sCoding
LFM2.5 24B A2B4-bit44K160 tok/sAgents
Mistral Small 3.2 24B3-bit20K27 tok/sChat, Coding, Agents
Ministral 3 14B6-bit20K28 tok/sCoding, Agents, Docs
Gemma 4 E4B8-bit128K53 tok/sChat, Coding, Docs
gpt-oss 20BMXFP4120K104 tok/sAgents
Ministral 3 8B8-bit40K34 tok/sDocs
Gemma 4 E2B8-bit128K96 tok/sChat
Nemotron 3 Nano 4B8-bit256K74 tok/sDocs
Llama 3.1 8B8-bit48K38 tok/sChat, Coding, Agents, Docs
Qwen3 0.6B8-bit40K295 tok/sNot rated
Qwen3 14B6-bit16K26 tok/sNot rated
Qwen3 VL 8B Instruct8-bit40K35 tok/sText and images
Qwen3 8B8-bit40K37 tok/sNot rated
Qwen2.5 7B Instruct8-bit32K40 tok/sNot rated
Qwen2.5 VL 7B Instruct8-bit112K37 tok/sText and images
Unlimited OCR8-bit32K218 tok/sReading documents
Qwen3 VL 4B Instruct8-bit64K62 tok/sText and images
Qwen3 1.7B8-bit40K129 tok/sNot rated
Qwen3 4B8-bit40K68 tok/sNot rated
Qwen3 4B Instruct 25078-bit72K68 tok/sNot rated
Qwen2.5 3B Instruct8-bit32K87 tok/sNot rated
Qwen3.6 27B3-bit20K23 tok/sText and images
DeepSeek OCR8-bit8K218 tok/sReading documents
Qwen2.5 Coder 14B Instruct6-bit16K26 tok/sCoding
Qwen2.5 VL 3B Instruct8-bit125K72 tok/sText and images
Qwen3.5 27B3-bit20K23 tok/sText and images
Qwen2.5 Coder 7B Instruct8-bit32K40 tok/sCoding
Mistral 7B Instruct V0.28-bit32K40 tok/sNot rated
Chandra OCR 28-bit256K52 tok/sReading documents
Qwen3 VL 2B Instruct8-bit112K124 tok/sText and images
MiniCPM5 2B8-bit128K113 tok/sNot rated
Gemma 3 4B IT8-bit128K75 tok/sText and images
Llama 3.2 3B Instruct8-bit96K85 tok/sNot rated
Qwen2.5 14B Instruct6-bit16K26 tok/sNot rated
Qwen3 1.7B Base8-bit32K150 tok/sNot rated
Diffusiongemma 26B A4B IT3-bit112K124 tok/sText and images
HunyuanOCR8-bit128K222 tok/sText and images
Qwen3 0.6B Base8-bit32K350 tok/sNot rated
Qwen2 VL 7B Instruct8-bit32K37 tok/sText and images
Ornith 1.5 9B8-bit144K33 tok/sNot rated
Unlimited OCR 28-bit128K217 tok/sReading documents
Granite 4.1 3B8-bit128K79 tok/sNot rated
Mistral 7B V0.18-bit32K40 tok/sNot rated
Qwen2.5 Coder 3B Instruct8-bit32K87 tok/sCoding
Meta Llama 3 8B Instruct Abliterated V38-bit8K38 tok/sNot rated
EXAONE 3.5 7.8B Instruct8-bit32K38 tok/sNot rated
Saiga Llama3 8B8-bit8K38 tok/sNot rated
DeepSeek R1 0528 Qwen3 8B8-bit44K37 tok/sNot rated
Gemma 3 12B IT8-bit28K25 tok/sText and images
Phi 3.5 Mini Instruct8-bit28K74 tok/sNot rated
Qwen3 4B Thinking 25078-bit72K68 tok/sNot rated
Phi 4 Mini Instruct8-bit80K72 tok/sNot rated
Dots.ocr8-bit128K97 tok/sReading documents

Fits, context and speed on the RTX 5060 Ti 16 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 247 are shown.

Image and video models

Image and video models that fit 16 GB
ModelBest fitPer outputStrongest at
LTX-2.55-bit2.6 minVideo
Z-Image Turbo16-bit20 sImages
FLUX.1 [dev]8-bit39 sImages
Qwen-Image5-bit14 minImages
Qwen-Image 2.116-bit5.2 minImages
FLUX.3 Klein16-bit29 sImages
Wan2.2 I2V A14B4-bit2.1 hVideo
FLUX.2 [klein] 9B8-bit9.7 sImages
Wan2.1 I2V 14B4-bit2.1 hVideo
Wan 2.1 T2V-14B4-bit2.6 hVideo
Krea 2 Turbo8-bit38 sImages
Stable Diffusion 3.5 Large8-bit2.7 minImages
Qwen Image 25125-bit14 minImages
MOSS-Image16-bit3.4 minImages
Wan 2.2 TI2V-5B16-bit29 minVideo
MOSS Image Base16-bit3.4 minImages
Wan 2.2 A14B4-bit2.1 hVideo
HunyuanVideo 1.58-bit2.2 hVideo

Machines with 16 GB

Questions

What is the best LLM for 16 GB of VRAM?

Of the open models we track, the strongest that run well on 16 GB are Qwen3.8 27B, Muse Glimmer 30B, Gemma 4 26B A4B. 247 text models fit in all.

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