What Runs on 6GB VRAM? Our Tested List of Local Models (Updated)
Language, image and world models we've actually run on one RTX 3060 Laptop 6GB with 64GB RAM. Every row links to the full test, and only models we ran ourselves are listed.
Test machine
Lenovo Legion Y9000P 2022: RTX 3060 Laptop 6GB, i7-12700H, 64GB DDR5-5200 dual channel, Windows 11 + WSL2 Ubuntu 24.04. Everything below depends on 64GB of RAM, which matters more than the GPU.
Language models
| Model | Quant | Usable? | Key numbers | Post |
|---|---|---|---|---|
| Qwen3.6-35B-A3B | UD-Q4_K_XL | Yes | ~21–23 t/s generation, 2.86 GB VRAM (experts in RAM) | 35B MoE |
| LFM2.5-24B-A2B | Q4_0 | Yes | 29.5 t/s at -ncmoe 30; all three tasks pass |
24B |
| Qwen3.6 APEX + MTP | — | Yes | MTP: +39% median generation speed, +688 MiB VRAM | APEX+MTP |
| Spark-X2.5-4B | — | Yes | Fits fully in VRAM; dozens of times faster than 35B on main tasks, but passes fewer of them | Spark vs 35B |
As a coding agent: LFM2.5-24B + pi passed 12/12, Claude Code 7/12. See the agent comparison.
Image / video
| Model | Usable? | Key numbers | Post |
|---|---|---|---|
| Qwen-Image 2.1 (INT8 + CPU offload) | Yes, slowly | ~3 min per 1024² image, ~4.8 GiB VRAM peak | Qwen-Image on 6GB, poetry posters |
| LingBot-World 1.3B | Produces clips, very slowly | 13 frames (0.8 s) in ~12 min, 5986 MiB VRAM peak | World model, Windows deploy |
Three lessons
- For MoE models, RAM matters more than VRAM. Keep the experts in RAM and the GPU needs only 2–3 GB. See -ncmoe tuning.
- Large dense models are mostly out. Dense weights that don't fit in 6GB have to stream over PCIe. This is inferred from the 35B run, which managed only 5 t/s with its experts on the GPU. We haven't tested a dense model directly.
- Image and video models work, measured in minutes. Fine for batch jobs, not for interactive use.
Related
FAQ
Can 6GB VRAM run a 35B model? The MoE Qwen3.6-35B-A3B can, with experts in 64GB RAM: ~21–23 t/s using 2.86 GB VRAM.