Best local LLMs for 20 GB of VRAM
Every open model we track that runs fully in 20 GB of GPU memory, biggest first. Checks are at Q4_K_M with about 4k context.
- Run fully at Q4
- 404 of 656
- Run at Q8 too
- 350
- Biggest at Q4
- Voxtral-Small-24B-2507
With 20 GB, the biggest model you can run comfortably is Voxtral-Small-24B-2507 (~17.9 GB at Q4). If you want higher quality at Q8, OLMo-2-1124-13B-Instruct is the biggest that fits.
Models that fit in 20 GB
Bar: Q4 memory needed vs. 20 GB. The tick marks 90%, our comfort line.FitsTight
| Model | Q4 need vs. 20 GB | Q4 / Q8 |
|---|---|---|
| Mistral-Small-3.2-24B-Instruct-250624.0B params, mistralai | 17.7 / 20.0 GB | Q4FitsQ8Won’t fit |
| LFM2.5-24B-A2B24.0B params, LiquidAI | 17.7 / 20.0 GB | Q4FitsQ8Won’t fit |
| gpt-oss-20b20.9B params, openai | 15.6 / 20.0 GB | Q4FitsQ8Won’t fit |
| Carbon-X (DNA)14.8B params, HuggingFaceBio | 11.3 / 20.0 GB | Q4FitsQ8Tight |
| Qwen2.5-Coder-14B-Instruct14.8B params, Qwen | 11.3 / 20.0 GB | Q4FitsQ8Tight |
| DeepSeek-R1-Distill-Qwen-14B14.8B params, deepseek-ai | 11.3 / 20.0 GB | Q4FitsQ8Tight |
| Qwen3-14B14.8B params, Qwen | 11.3 / 20.0 GB | Q4FitsQ8Tight |
| phi-414.7B params, microsoft | 11.2 / 20.0 GB | Q4FitsQ8Tight |
| OLMo-2-1124-13B-Instruct13.7B params, allenai | 10.6 / 20.0 GB | Q4FitsQ8Fits |
| Mistral-Nemo-Instruct-240712.2B params, mistralai | 9.5 / 20.0 GB | Q4FitsQ8Fits |
| gemma-3-12b-it12.2B params, google | 9.5 / 20.0 GB | Q4FitsQ8Fits |
| Mellum2.1-12B-A2.5B-Thinking12.1B params, JetBrains | 9.5 / 20.0 GB | Q4FitsQ8Fits |
| Gemma 4 12B12.0B params, google | 9.3 / 20.0 GB | Q4FitsQ8Fits |
| Ornith-1.5-9B9.7B params, ornith-ai | 7.7 / 20.0 GB | Q4FitsQ8Fits |
| Qwen3.5-9B9.7B params, Qwen | 7.7 / 20.0 GB | Q4FitsQ8Fits |
| Qwen2.5-VL-7B-Instruct8.3B params, Qwen | 6.8 / 20.0 GB | Q4FitsQ8Fits |
| Qwen3-8B8.2B params, Qwen | 6.7 / 20.0 GB | Q4FitsQ8Fits |
| granite-3.3-8b-instruct8.2B params, ibm-granite | 6.7 / 20.0 GB | Q4FitsQ8Fits |
| Llama-3.1-8B-Instruct8.0B params, meta-llama | 6.6 / 20.0 GB | Q4FitsQ8Fits |
| Nemotron 3.5 Bolt E4B8.0B params, nvidia | 6.6 / 20.0 GB | Q4FitsQ8Fits |
| Gemma 4 E4B8.0B params, google | 6.6 / 20.0 GB | Q4FitsQ8Fits |
| Qwen2.5-Coder-7B-Instruct7.6B params, Qwen | 6.3 / 20.0 GB | Q4FitsQ8Fits |
| DeepSeek-R1-Distill-Qwen-7B7.6B params, deepseek-ai | 6.3 / 20.0 GB | Q4FitsQ8Fits |
| Qwen2.5-7B-Instruct7.6B params, Qwen | 6.3 / 20.0 GB | Q4FitsQ8Fits |
| Mistral-7B-Instruct-v0.37.2B params, mistralai | 6.1 / 20.0 GB | Q4FitsQ8Fits |
| Gemma 4 E2B5.1B params, google | 4.6 / 20.0 GB | Q4FitsQ8Fits |
| Qwen3.5-4B4.7B params, Qwen | 4.2 / 20.0 GB | Q4FitsQ8Fits |
| gemma-3-4b-it4.3B params, google | 4.0 / 20.0 GB | Q4FitsQ8Fits |
| Spark-X2.6-4B4.1B params, XHToken | 3.9 / 20.0 GB | Q4FitsQ8Fits |
| Qwen3-4B4.0B params, Qwen | 3.8 / 20.0 GB | Q4FitsQ8Fits |
| Nemotron Labs Dory 1.0 4B4.0B params, nvidia | 3.8 / 20.0 GB | Q4FitsQ8Fits |
| Phi-4-mini-instruct3.8B params, microsoft | 3.7 / 20.0 GB | Q4FitsQ8Fits |
| Llama-3.2-3B-Instruct3.2B params, meta-llama | 3.2 / 20.0 GB | Q4FitsQ8Fits |
| MiniCPM5-2B2.5B params, openbmb | 2.8 / 20.0 GB | Q4FitsQ8Fits |
| Qwen3-1.7B2.0B params, Qwen | 2.4 / 20.0 GB | Q4FitsQ8Fits |
| LFM3-1.7B1.7B params, LiquidAI | 2.2 / 20.0 GB | Q4FitsQ8Fits |
| Llama-3.2-1B-Instruct1.2B params, meta-llama | 1.9 / 20.0 GB | Q4FitsQ8Fits |
| gemma-3-1b-it1.0B params, google | 1.7 / 20.0 GB | Q4FitsQ8Fits |
| Qwen3-0.6B0.8B params, Qwen | 1.5 / 20.0 GB | Q4FitsQ8Fits |
| Gemma 4 26B A4B25.8B params, google | 19.0 / 20.0 GB | Q4TightQ8Won’t fit |
Hardware we track with 20–23 GB usable
- Radeon RX 7900 XT20 GB usable, $899
- NVIDIA RTX 4000 Ada 20GB20 GB usable, $1250
Quick answers
- What is the biggest LLM I can run with 20 GB of VRAM?
- Voxtral-Small-24B-2507 (24.3B parameters) is the biggest model we track that runs comfortably in 20 GB at Q4_K_M, needing about 17.9 GB with a ~4k context. At Q8 the biggest is OLMo-2-1124-13B-Instruct.
- How many models fit in 20 GB of VRAM?
- 404 of the 656 open models we track run fully in 20 GB at Q4 with room to spare, 350 also fit at Q8, and 9 more fit only with a short context.
- Can I run a model that needs more than 20 GB?
- Yes, with llama.cpp or Ollama you can offload some layers to system RAM, but speed drops sharply. For full-speed local inference the whole model plus its context cache should fit in GPU memory; we use 90% of 20 GB as the comfort line.
- Which hardware has about 20 GB of usable memory?
- The cheapest current device we track in this range is the Radeon RX 7900 XT (20 GB usable, $899 list price as of our last snapshot). 2 tracked devices have 20–23 GB usable.
Other memory sizes
4 GB · 6 GB · 8 GB · 10 GB · 12 GB · 16 GB · 24 GB · 32 GB · 48 GB · 64 GB · 96 GB · 128 GB