Best open models for an 8 GB GPU
An 8 GB graphics card like the RTX 3060 Ti, RTX 4060, or a laptop GPU can run smaller language models and vision-language models. These models need about 3–8 GB of memory to load, leaving room for context and generation.
Last updated 3 November 2026 · refreshed after every Hugging Face sync · 15 models
| # | Model | Size | Memory | License | Downloads (30 days) | Likes |
|---|---|---|---|---|---|---|
| 1 | Qwen2.5-1.5B-Instruct Qwen | 1.5B · FP16 | ~4 GB | Apache 2.0 | 4.6M | 926 |
| 2 | Ornith-1.5-9B-GGUF ornith-ai | 9B · GGUF | ~6 GB | MIT | 5.4M | 625 |
| 3 | Qwen3-1.7B Qwen | 1.7B · FP16 | ~5 GB | Apache 2.0 | 4.2M | 611 |
| 4 | Llama-3.2-1B-Instruct meta-llama | 1B · FP16 | ~3 GB | Llama 3.2 license | 1.8M | 1.9K |
| 5 | Qwen2.5-3B-Instruct Qwen | 3B · FP16 | ~8 GB | Custom (other) | 3.1M | 631 |
| 6 | Ornith-1.0-9B-GGUF ornith-ai | 9B · GGUF | ~6 GB | MIT | 2.4M | 699 |
| 7 | Llama-3.2-3B-Instruct meta-llama | 3B · FP16 | ~8 GB | Llama 3.2 license | 995.2K | 2.9K |
| 8 | MiniCPM5-2B openbmb | 2B · FP16 | ~5 GB | Apache 2.0 | 1.1M | 1.8K |
| 9 | TinyLlama-1.1B-Chat-v1.0 TinyLlama | 1.1B · FP16 | ~3 GB | Apache 2.0 | 929.6K | 1.9K |
| 10 | DeepSeek-R1-Distill-Qwen-1.5B deepseek-ai | 1.5B · FP16 | ~4 GB | MIT | 784.6K | 1.7K |
| 11 | gemma-3-1b-it | 1B · FP16 | ~3 GB | Gemma license | 826.6K | 1.3K |
| 12 | Llama-3.2-1B meta-llama | 1B · FP16 | ~3 GB | Llama 3.2 license | 479.7K | 2.7K |
| 13 | LocateAnything-3B nvidia | 3B · FP16 | ~8 GB | Custom (other) | 322.5K | 3.2K |
| 14 | Qwen3-8B-AWQ Qwen | 8B · AWQ | ~5 GB | Apache 2.0 | 2.1M | 70 |
| 15 | Qwen2.5-7B-Instruct-AWQ Qwen | 7B · AWQ | ~5 GB | Apache 2.0 | 1.8M | 61 |
How we rank
Language and vision-language models whose published weights need about 3–8 GB. We include models seen on Hugging Face in the last 14 days with at least 10,000 downloads in the last 30 days. Score = log10(30-day downloads) + ½ × log10(likes + 1), so popularity leads and likes break near-ties. We keep one entry per model name and leave out adult and uncensored builds. Memory is estimated from parameter count and the published file format (16-bit ≈ 2 bytes per parameter, 4-bit ≈ 0.5) plus 20% headroom; long contexts need more. A quantised copy of a bigger model may also fit — check the model page.