Best open models for a 12–16 GB GPU
A 12–16 GB graphics card like the RTX 3060 12 GB, RTX 4060 Ti 16 GB, RTX 4070 Ti Super, RTX 5060 Ti 16 GB, or A4000 can run mid-sized language models and vision-language models. These models need about 8–16 GB of memory to load.
Last updated 3 November 2026 · refreshed after every Hugging Face sync · 15 models
| # | Model | Size | Memory | License | Downloads (30 days) | Likes |
|---|---|---|---|---|---|---|
| 1 | Qwen3-4B Qwen | 4B · FP16 | ~10 GB | Apache 2.0 | 8.4M | 780 |
| 2 | Qwen3.5-4B Qwen | 4B · FP16 | ~10 GB | Apache 2.0 | 6.4M | 1.2K |
| 3 | Qwen3-VL-4B-Instruct Qwen | 4B · FP16 | ~10 GB | Apache 2.0 | 3.9M | 523 |
| 4 | Qwen3-4B-Instruct-2507 Qwen | 4B · FP16 | ~10 GB | Apache 2.0 | 2.6M | 1K |
| 5 | Qwen2.5-3B-Instruct Qwen | 3B · FP16 | ~8 GB | Custom (other) | 3.1M | 631 |
| 6 | Llama-3.2-3B-Instruct meta-llama | 3B · FP16 | ~8 GB | Llama 3.2 license | 995.2K | 2.9K |
| 7 | NVIDIA-Nemotron-3-Nano-4B-BF16 nvidia | 4B · FP16 | ~10 GB | Custom (other) | 1.9M | 142 |
| 8 | Qwen3-14B-AWQ Qwen | 14B · AWQ | ~9 GB | Apache 2.0 | 2M | 89 |
| 9 | LocateAnything-3B nvidia | 3B · FP16 | ~8 GB | Custom (other) | 322.5K | 3.2K |
| 10 | Qwen3-4B-Instruct-2507-FP8 Qwen | 4B · FP16 | ~10 GB | Apache 2.0 | 1.3M | 91 |
| 11 | Qwen2.5-14B-Instruct-AWQ Qwen | 14B · AWQ | ~9 GB | Apache 2.0 | 1.2M | 53 |
| 12 | gemma-4-12B-coder-fable5-composer2.5-v1-GGUF yuxinlu1 | 12B · GGUF | ~8 GB | Apache 2.0 | 110.7K | 3K |
| 13 | OTel-2.0-Decision-26B-A4B-4bit farbodtavakkoli | 26B · AWQ | ~16 GB | Apache 2.0 | 2.2M | 6 |
| 14 | LFM2.5-24B-A2B-GGUF LiquidAI | 24B · GGUF | ~15 GB | Custom (other) | 148.6K | 235 |
| 15 | Mellum2.1-12B-A2.5B-Thinking-GGUF JetBrains | 12B · GGUF | ~8 GB | Apache 2.0 | 152K | 132 |
How we rank
Language and vision-language models whose published weights need about 8–16 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.