selfhost.directory

The index of self-hostable projects — ranked, categorized and served with live updates.
Own your data and stop giving it to big tech.

Best LLMs and AI models for 16 GB VRAM (2026)

2 439 open-weight models fit fully in 16 GB of GPU memory at their recommended quantisation. Each section puts the strongest model that still fits comfortably first, with the best quantisation and an estimated generation speed.

A GPU with 16 GB of VRAM leaves 16 GB usable for model weights, enough to run several open-weight language models locally. Models in the 8B to 12B range fit, and so does gpt-oss-20b at 20.9B parameters. Popular picks include Llama-3.1-8B-Instruct from Meta, Qwen3-8B from Qwen, and gemma-4-12B-it from Google.

Best chat models for 16 GB

All 921 that fit →
#ModelParamsBest quantMemorySpeedLicence
120.9BA4.2BQ8_0≈13 GBFits in memory~55 tok/sCommercial OK
214.8BQ6_K≈14 GBFits in memory~20 tok/sCommercial OK
38.2BQ8_0≈10 GBFits in memory~28 tok/sCommercial OK
414.8BQ6_K≈14 GBFits in memory~20 tok/sCommercial OK
5
microsoft logophi-4microsoft
14.7BQ6_K≈14 GBFits in memory~20 tok/sCommercial OK
6
deepgrove logomaple-previewdeepgrove
20.2BA1.5BQ5_K_M≈15 GBFits in memory~232 tok/sCommercial OK
7
ornith-ai logoOrnith-1.5-9Bornith-ai
9.7BQ8_0≈11 GBFits in memory~24 tok/sCommercial OK
8
inclusionAI logoLing-3.0-tinyinclusionAI
7.9BA1.4BQ8_0≈10 GBFits in memory~168 tok/sCommercial OK

Best code models for 16 GB

All 79 that fit →
#ModelParamsBest quantMemorySpeedLicence
112.1BA2.5BQ8_0≈14 GBFits in memory~93 tok/sCommercial OK
214.8BQ6_K≈14 GBFits in memory~20 tok/sCommercial OK
312.1BA2.5BQ8_0≈14 GBFits in memory~93 tok/sCommercial OK
415.7BA2.7BQ5_K_M≈15 GBFits in memory~130 tok/sConditions
516BQ6_K≈14 GBFits in memory~19 tok/sConditions
6
Tesslate logoOmniCoder-9BTesslate
9.4BQ8_0≈11 GBFits in memory~25 tok/sCommercial OK
7
bigcode logostarcoderbigcode
15.8BQ6_K≈15 GBFits in memory~19 tok/sConditions
87.6BQ8_0≈8.9 GBFits in memory~30 tok/sCommercial OK

Best reasoning models for 16 GB

All 51 that fit →
#ModelParamsBest quantMemorySpeedLicence
114.8BQ6_K≈14 GBFits in memory~20 tok/sCommercial OK
214.7BQ6_K≈14 GBFits in memory~20 tok/sCommercial OK
314.7BQ6_K≈14 GBFits in memory~20 tok/sCommercial OK
415.9BA3BQ5_K_M≈14 GBFits in memory~115 tok/sNon-commercial
521.8BA3BQ4_K_M≈14 GBFits in memory~135 tok/sCommercial OK
614.8BQ6_K≈14 GBFits in memory~20 tok/sCommercial OK
77.3BQ8_0≈12 GBFits in memory~32 tok/sCommercial OK
8
inclusionAI logoRing-mini-2.0inclusionAI
16.3BA1.4BQ6_K≈14 GBFits in memory~209 tok/sCommercial OK

Best vision models for 16 GB

All 336 that fit →
#ModelParamsBest quantMemorySpeedLicence
19.7BQ8_0≈11 GBFits in memory~24 tok/sCommercial OK
219.4BA3BQ4_K_M≈13 GBFits in memory~135 tok/sCommercial OK
316.4BA3BQ5_K_M≈15 GBFits in memory~115 tok/sCommercial OK
413.2BQ6_K≈13 GBFits in memory~23 tok/sConditions
5
Cloudflare logoclef-flashCloudflare
9.4BQ8_0≈11 GBFits in memory~25 tok/sCommercial OK
616.4BA3BQ5_K_M≈14 GBFits in memory~115 tok/sCommercial OK
78.8BQ8_0≈10 GBFits in memory~26 tok/sCommercial OK
812.2BQ6_K≈13 GBFits in memory~25 tok/sConditions

Best image gen models for 16 GB

All 112 that fit →
#ModelParamsBest quantMemoryLicence
17.1BQ8_0≈8.9 GBFits in memoryNon-commercial
27.1BQ8_0≈8.9 GBFits in memoryNon-commercial
3
Tongyi-MAI logoZ-Image-TurboTongyi-MAI
6.2BFP16 / BF16≈14 GBFits in memoryCommercial OK
4
OpenMOSS-Team logoMOSS-Image-SFTOpenMOSS-Team
5.1BFP16 / BF16≈12 GBFits in memoryCommercial OK
512.8BQ4_K_M≈9.2 GBFits in memoryConditions
6
OpenMOSS-Team logoMOSS-Image-BaseOpenMOSS-Team
5.1BFP16 / BF16≈12 GBFits in memoryCommercial OK
7
Tongyi-MAI logoZ-ImageTongyi-MAI
6.2BFP16 / BF16≈14 GBFits in memoryCommercial OK
812.8BQ4_K_M≈9.2 GBFits in memoryConditions

Best STT models for 16 GB

All 100 that fit →
#ModelParamsBest quantMemoryLicence
12.1BFP16 / BF16≈5.1 GBFits in memoryCommercial OK
22.3BFP16 / BF16≈5.8 GBFits in memoryCommercial OK
38.7BQ8_0≈11 GBFits in memoryCommercial OK
4
microsoft logoVibeVoice-ASRmicrosoft
8.7BQ8_0≈11 GBFits in memoryCommercial OK
54.4BFP16 / BF16≈10 GBFits in memoryCommercial OK
62.3BFP16 / BF16≈5.7 GBFits in memoryCommercial OK
7627MFP16 / BF16≈1.9 GBFits in memoryCommercial OK
8909MFP16 / BF16≈2.5 GBFits in memoryCommercial OK

Best TTS models for 16 GB

All 111 that fit →
#ModelParamsBest quantMemoryLicence
1
m-a-p logoYuE2-3Bm-a-p
3.6BFP16 / BF16≈8.6 GBFits in memoryNon-commercial
2
BreezeBlue logoBreeze-TTS-2BreezeBlue
3.5BFP16 / BF16≈8.3 GBFits in memoryNon-commercial
3
openbmb logoVoxCPM2openbmb
2.3BFP16 / BF16≈5.6 GBFits in memoryCommercial OK
4
k2-fsa logoOmniVoicek2-fsa
613MFP16 / BF16≈1.9 GBFits in memoryCheck licence
51.9BFP16 / BF16≈4.8 GBFits in memoryCommercial OK
6915MFP16 / BF16≈2.5 GBFits in memoryCommercial OK
71.5BFP16 / BF16≈3.8 GBFits in memoryCommercial OK
8
fishaudio logos2-profishaudio
4.6BFP16 / BF16≈11 GBFits in memoryNon-commercial

Best embedding models for 16 GB

All 241 that fit →
#ModelParamsBest quantMemoryLicence
1744MFP16 / BF16≈2.2 GBFits in memoryCommercial OK
27.6BQ8_0≈9.5 GBFits in memoryCommercial OK
3303MFP16 / BF16≈1.2 GBFits in memoryConditions
4596MFP16 / BF16≈1.8 GBFits in memoryCommercial OK
54BFP16 / BF16≈9.5 GBFits in memoryCommercial OK
68BQ8_0≈9.9 GBFits in memoryCommercial OK
71.1BFP16 / BF16≈3.0 GBFits in memoryCommercial OK
88.1BQ8_0≈10 GBFits in memoryCommercial OK

GPUs with 16–18 GB 57

Frequently asked questions

What is the best LLM for 16 GB of VRAM?
Our top chat pick for 16 GB is gpt-oss-20b (openai) at Q8_0, needing about 12.72 GB, at roughly 55 tokens/s. The ranking favours the strongest model that still fits comfortably, then benchmarks, popularity and release date.
What is the best coding model for 16 GB of VRAM?
Mellum2.1-12B-A2.5B-Thinking (JetBrains) at Q8_0, needing about 13.58 GB, at roughly 93 tokens/s.
How big a model fits in 16 GB of VRAM?
Dense LLMs up to about 22.1B parameters fit fully in 16 GB at Q4_K_M or better, with an 8K-token context. Longer contexts need more memory for the KV cache.
How many open-weight models run on 16 GB of VRAM?
2439 open-weight models on this site fit fully in 16 GB at their recommended quantisation, across 19 categories (chat, code, vision, image generation, speech and more).
How fast are models on 16 GB of VRAM?
Token generation is limited by memory bandwidth. The speeds on this page assume about 448 GB/s, typical for 16 GB cards; a faster or slower card scales roughly in proportion. Open your GPU’s page for estimates with its own bandwidth.
Which GPUs have 16 GB of VRAM?
Among others: GeForce RTX 5060 Ti 16GB, GeForce RTX 5070 Ti, GeForce RTX 5080, Radeon RX 9060 XT 16GB, Radeon RX 9070 XT, Apple M4 (16 GB). Each GPU page lists what fits and how fast it runs.
What if a model doesn’t fit in 16 GB?
llama.cpp, Ollama and LM Studio can offload part of an LLM to system RAM. It runs, but generation slows down a lot because RAM is far slower than VRAM. The 20 GB page shows what one more step of memory buys.

Memory figures are estimates: weights at the given quantisation plus an 8K-token KV cache and runtime overhead. On Macs and other unified-memory machines only part of the memory is available to the GPU, see the Apple Silicon pages.