Local Model Advisor
Which open-weight models fit on your hardware? Enter the memory your model will run in: GPU VRAM on a discrete card, or unified memory on Apple Silicon. Parameter counts come from each model's Hugging Face repository; memory needs are estimated from them with the formula below.
Example: 54 models fit in 24 GB (an RTX 4090)
144 of 1,345 open-weight models in the catalog have a parameter count on Hugging Face; the rest are not shown rather than guessed. Sorted by the highest precision that fits, then by model size. Pick your hardware above to see your own results.
| Model | Parameters | FP16 / BF16 | 8-bit | 4-bit |
|---|---|---|---|---|
| Qwen3.5-9B Qwen · Hugging Face | 9.7B | ~23.2 GB | ~12.4 GB | ~7 GB |
| Ministral 3 8B Mistral · Hugging Face | 8.9B | ~21.5 GB | ~11.4 GB | ~6.5 GB |
| Ministral 3 8B 2512 (batch) Mistral · Hugging Face | 8.9B | ~21.5 GB | ~11.4 GB | ~6.5 GB |
| Granite 4.2 8B IBM · Hugging Face | 8.8B | ~21.1 GB | ~11.3 GB | ~6.4 GB |
| Qwen3 VL 8B Instruct Qwen · Hugging Face | 8.8B | ~21.1 GB | ~11.2 GB | ~6.4 GB |
| UI-TARS 7B Bytedance · Hugging Face | 8.3B | ~20 GB | ~10.6 GB | ~6 GB |
| Llama 3.1 8B Instruct Meta · Hugging Face | 8.0B | ~19.3 GB | ~10.3 GB | ~5.8 GB |
| Llama 3 8B Lunaris Sao10k · Hugging Face | 8.0B | ~19.3 GB | ~10.3 GB | ~5.8 GB |
| Hy-MT2-7B Tencent · Hugging Face | 8.0B | ~19.3 GB | ~10.3 GB | ~5.8 GB |
| Qwen2.5 7B Instruct Qwen · Hugging Face | 7.6B | ~18.3 GB | ~9.8 GB | ~5.5 GB |
| Reka Edge Rekaai · Hugging Face | 7.1B | ~17.2 GB | ~9.1 GB | ~5.2 GB |
| Gemma 3 4B Google · Hugging Face | 4.3B | ~10.4 GB | ~5.5 GB | ~3.1 GB |
| Nemotron 3.5 Content Safety NVIDIA · Hugging Face | 4.3B | ~10.4 GB | ~5.5 GB | ~3.1 GB |
| Ministral 3 3B Mistral · Hugging Face | 3.8B | ~9.3 GB | ~5 GB | ~2.8 GB |
| Llama 3.2 3B Instruct Meta · Hugging Face | 3.2B | ~7.8 GB | ~4.1 GB | ~2.4 GB |
| Granite 4.0 Micro IBM · Hugging Face | 3.2B | ~7.7 GB | ~4.1 GB | ~2.3 GB |
| Hy-MT2-1.8B Tencent · Hugging Face | 2.0B | ~4.9 GB | ~2.6 GB | ~1.5 GB |
| Llama 3.2 1B Instruct Meta · Hugging Face | 1.2B | ~3 GB | ~1.6 GB | ~0.9 GB |
| Qwen3 14B Qwen · Hugging Face | 15B | ~35.5 GB | ~18.9 GB | ~10.7 GB |
| Phi 4 Microsoft · Hugging Face | 15B | ~35.2 GB | ~18.7 GB | ~10.6 GB |
| Ministral 3 14B Mistral · Hugging Face | 14B | ~33.5 GB | ~17.8 GB | ~10.1 GB |
| Mistral Nemo Mistral · Hugging Face | 12B | ~29.4 GB | ~15.7 GB | ~8.9 GB |
| UnslopNemo 12B Thedrummer · Hugging Face | 12B | ~29.4 GB | ~15.7 GB | ~8.9 GB |
| Gemma 3 12B Google · Hugging Face | 12B | ~29.3 GB | ~15.6 GB | ~8.8 GB |
Showing the top 24 of 54.
Memory ≈ parameters × bits per weight ÷ 8, × 1.2 for runtime buffers and a KV cache of a few thousand tokens. Bits per weight: FP16/BF16 = 16, 8-bit ≈ 8.5, 4-bit ≈ 4.8 (typical of GGUF Q8_0 and Q4_K_M).
Long contexts need extra memory for the KV cache, and mixture-of-experts models still need all their weights loaded even though only some are active per token. Treat these figures as a first filter, then check the model card and your runtime (Ollama, llama.cpp, MLX or vLLM) before downloading.