LM Studio: which models fit your hardware
LM Studio downloads GGUF files from Hugging Face (and MLX files on Apple Silicon) and runs them with llama.cpp. So the question "will it fit?" comes down to file size plus context. Pick your hardware to see which of the 60 most-downloaded open models load fully onto it, and at which quant.
Loading…
Best quant that fits fully in GPU memory (what LM Studio calls full GPU offload), using real GGUF file sizes where we have them. When nothing fits, LM Studio can still run the model with partial offload to the CPU, much more slowly; the llama.cpp offload guide estimates how fast.
How LM Studio's quants map to memory
In LM Studio's download panel each model comes in several files, named by quantization. The name says roughly how many bits each weight uses. Fewer bits means a smaller file and slightly lower quality:
| File name contains | Bits per weight | Quality | Llama 3.1 8B Instruct | Qwen3 14B | gpt-oss 20B |
|---|---|---|---|---|---|
Q2_K | 3 | rough | 3.2 GB file 4.6 GB at 8K | ~7.1 GB at 8K | 11.6 GB file 11.7 GB at 8K |
Q3_K_M | 3.91 | usable | 4.0 GB file 5.4 GB at 8K | ~8.7 GB at 8K | 11.6 GB file 11.7 GB at 8K |
Q4_K_M | 4.89 | good, most popular | 4.9 GB file 6.2 GB at 8K | 9.0 GB file 10.3 GB at 8K | 11.7 GB file 11.8 GB at 8K |
Q5_K_M | 5.7 | very good | 5.7 GB file 7.0 GB at 8K | 10.5 GB file 11.7 GB at 8K | 11.7 GB file 11.9 GB at 8K |
Q6_K | 6.56 | near lossless | 6.6 GB file 7.8 GB at 8K | 12.1 GB file 13.2 GB at 8K | 12.0 GB file 12.2 GB at 8K |
Q8_0 | 8.5 | effectively lossless | 8.5 GB file 9.6 GB at 8K | 15.7 GB file 16.6 GB at 8K | 12.1 GB file 12.3 GB at 8K |
F16 / BF16 | 16 | original weights | ~16.6 GB at 8K | ~29.4 GB at 8K | 13.8 GB file 13.8 GB at 8K |
File sizes in decimal GB from the GGUF repos (bartowski/Meta-Llama-3.1-8B-Instruct-GGUF, Qwen/Qwen3-14B-GGUF, bartowski/openai_gpt-oss-20b-GGUF); memory at 8K context in binary GB, including KV cache and overhead. "~" means no file of that quant in the repo, so the size is estimated from bits per weight. gpt-oss ships its expert weights in MXFP4, which is why every quant is about the same size.
Q4_K_M is the usual starting point: about a quarter of the original size with little quality loss. If a model fits with room to spare, Q5_K_M or Q6_K costs a little more memory for a little more quality. Files starting with IQ (such as IQ4_XS) are smaller alternatives at the same nominal bit width; they are not in our tables. On a Mac, LM Studio's MLX 4-bit files are close to Q4_K_M in size, but not identical.
LM Studio system requirements
| Mac | Apple Silicon (M1 or later), macOS 14 or newer. 16 GB of memory recommended; 8 GB Macs work with small models and short contexts. Intel Macs are not supported. |
| Windows | x64 with AVX2, or ARM (Snapdragon X Elite). At least 16 GB of RAM and 4 GB of dedicated VRAM recommended. |
| Linux | x64 or ARM64 AppImage; Ubuntu 20.04 or newer. |
| What really decides it | Your GPU memory, as above. A 16 GB GPU runs the 12–15B class at 4-bit; 24 GB runs most 20–32B models. See models by VRAM size. |
From LM Studio's system requirements, checked Nov 17, 2026.
Questions
Why does LM Studio say a model is "partial GPU offload"?
The file plus its context doesn't fit in your GPU memory, so some layers will run on the CPU from system RAM. It works, but expect a few tokens per second instead of dozens. A smaller quant or shorter context may make it fit fully; the table above shows the best quant that does.
LM Studio, Ollama or llama.cpp?
Same engine underneath, different front ends: LM Studio is a desktop app with a model browser, Ollama is a background service with a simple command line, and llama.cpp gives you every setting. The local LLM guide compares them.
Automated data from the Hugging Face Hub API and GGUF file listings; formulas on the methodology page. Not affiliated with LM Studio. Report an error.