Local model fit check
Which rig can run which open model.
Memory at Q4, Q8, and FP16, the machines that clear it, and a list price. Amazon links search for the product. They do not pretend to be tonight's street price.
- Models
- 49
- Machines
- 64
- Quants
- 3
Example · Llama 3.3 70B Instruct
42.5 GB at Q4
70.6B total weights. The figure is an estimate.
- GeForce RTX 4090Doesn't fit
- GeForce RTX 5090Doesn't fit
- Mac mini M4 Pro 64GBFits at Q4
- Framework Desktop 128GBFits at Q4
Popular models
A short list of models people actually try to run at home. The full catalog is on the models page.
- GemmaMoE
Gemma 4 26B-A4B
25.8B
- Q4
- 17.3 GB
- Q8
- 27.2 GB
- FP16
- 61.9 GB
- Qwen
Qwen3 8B
8.19B
- Q4
- 6.2 GB
- Q8
- 9.9 GB
- FP16
- 19.7 GB
- Gemma
Gemma 4 31B
31.3B
- Q4
- 19.7 GB
- Q8
- 34.0 GB
- FP16
- 75.1 GB
- Qwen
Qwen3.8 27B
27.8B
- Q4
- 17.0 GB
- Q8
- 29.6 GB
- FP16
- 66.7 GB
- gpt-ossMoE
gpt-oss-20b
20.9B
- Q4
- 11.8 GB
- Q8
- 12.3 GB
- FP16
- 50.2 GB
- Llama
Llama 3.2 1B Instruct
1.24B
- Q4
- 0.8 GB
- Q8
- 1.3 GB
- FP16
- 3.0 GB
- Llama
Llama 3.1 8B Instruct
8.03B
- Q4
- 4.9 GB
- Q8
- 9.6 GB
- FP16
- 19.3 GB
- QwenMoE
Qwen3.6 35B-A3B
36B
- Q4
- 22.3 GB
- Q8
- 37.1 GB
- FP16
- 86.3 GB
Popular checks
Static pages for pairs people look up. The full curated list is on the can-run index.
Best-value hardware
Ranked by how many catalog models fit at Q4, per $100 of list price. Bandwidth is not in the score, so a cheaper card can beat a faster one. Configure-to-order Macs without a cited price are left out.
- Radeon RX 7600 XT
16.0 GB · 288 GB/s · $329
21 models at Q4 · 6.38 per $100
15.2 GB usable after the 95% allowance.
- GeForce RTX 3060 12GB
12.0 GB · 360 GB/s · $329
20 models at Q4 · 6.08 per $100
11.4 GB usable after the 95% allowance.
- Radeon RX 9060 XT 16GB
16.0 GB · 320 GB/s · $349
21 models at Q4 · 6.02 per $100
15.2 GB usable after the 95% allowance.
- GeForce RTX 5050
8.0 GB · 320 GB/s · $249
14 models at Q4 · 5.62 per $100
7.6 GB usable after the 95% allowance.
- Radeon RX 7600
8.0 GB · 288 GB/s · $269
14 models at Q4 · 5.20 per $100
7.6 GB usable after the 95% allowance.
- GeForce RTX 5060 Ti 16GB
16.0 GB · 448 GB/s · $429
21 models at Q4 · 4.90 per $100
15.2 GB usable after the 95% allowance.
How the estimate works
Estimated memory is parameter count × bytes per weight × 1.20 when no GGUF file size is stored. Q4 uses 0.5 byte per weight, Q8 uses 1, and FP16 uses 2. The 20% is a modest allowance for KV cache and runtime overhead, not a budget for a 128K or million-token context. When a matching GGUF file is stored, the fit check uses that file size plus an FP16 KV-cache estimate at 8,192 tokens.
Full method, including KV cache and the tokens/sec roofline
Data updated . Last reviewed the same day, when this catalog snapshot was written.
Discrete GPUs keep 95% of VRAM. Macs keep 70% of unified memory, because macOS shares the pool. Strix Halo mini PCs keep 75%, about what the GPU can be assigned. Mixture-of-experts models are sized by total weights, not active parameters. A fit labeled Tight has less than 2 GB of headroom.