Best open models for a 24 GB GPU
A 24 GB graphics card like the RTX 3090, RTX 4090, A5000, or L4 can run larger language models and vision-language models. These models need about 12–24 GB of memory to load. That covers most 7–9B models at full 16-bit precision and many 20–30B models in 4-bit.
Last updated 3 November 2026 · refreshed after every Hugging Face sync · 15 models
| # | Model | Size | Memory | License | Downloads (30 days) | Likes |
|---|---|---|---|---|---|---|
| 1 | Llama-3.1-8B-Instruct meta-llama | 8B · FP16 | ~20 GB | Llama 3.1 license | 4.5M | 8.8K |
| 2 | Qwen3-8B Qwen | 8B · FP16 | ~20 GB | Apache 2.0 | 7.5M | 2.2K |
| 3 | Qwen3.5-9B Qwen | 9B · FP16 | ~22 GB | Apache 2.0 | 6.4M | 2.3K |
| 4 | Qwen2.5-7B-Instruct Qwen | 7B · FP16 | ~17 GB | Apache 2.0 | 5.9M | 2.5K |
| 5 | Qwen2.5-VL-7B-Instruct Qwen | 7B · FP16 | ~17 GB | Apache 2.0 | 5.9M | 1.8K |
| 6 | Qwen3-VL-8B-Instruct Qwen | 8B · FP16 | ~20 GB | Apache 2.0 | 6.9M | 1.2K |
| 7 | Ternary-Bonsai-2-27B-gguf prism-ml | 27B · GGUF | ~17 GB | Apache 2.0 | 1.8M | 2.9K |
| 8 | Mistral-7B-Instruct-v0.2 mistralai | 7B · FP16 | ~17 GB | Apache 2.0 | 1.7M | 3.3K |
| 9 | Ornith-1.5-35B-A3B-GGUF ornith-ai | 35B · GGUF | ~21 GB | MIT | 4M | 571 |
| 10 | Qwen3.8-27B-GSQ-RCO-GGUF ISTA-DASLab | 27B · GGUF | ~17 GB | Apache 2.0 | 1.3M | 2.5K |
| 11 | Ornith-1.0-35B-GGUF ornith-ai | 35B · GGUF | ~21 GB | MIT | 1.5M | 1.1K |
| 12 | Qwen2.5-Coder-7B-Instruct Qwen | 7B · FP16 | ~17 GB | Apache 2.0 | 1.4M | 859 |
| 13 | Qwen3.8-27B-MLX-4bit lmstudio-community | 27B · MLX | ~17 GB | Apache 2.0 | 3.6M | 116 |
| 14 | Qwen3-Coder-30B-A3B-Instruct-GGUF unsloth | 30B · GGUF | ~18 GB | Apache 2.0 | 1.1M | 1.2K |
| 15 | Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF ukisai | 27B · GGUF | ~17 GB | Custom (other) | 1.6M | 478 |
How we rank
Language and vision-language models whose published weights need about 12–24 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.