How to use this list
Find the section that matches your available VRAM or unified memory, at Q4 quantization (the most common default for local use, and what these tiers assume unless noted). If you're not sure what tier you're in, run your numbers through the calculator first.
8GB tier: entry GPU or a base laptop
| Model | Best for | Notes |
|---|---|---|
| Llama 3.1 8B / Llama 3.3 8B | General chat, writing | The most widely supported starting model; broad tool compatibility |
| Mistral 7B | Speed | Fastest of this group; slightly less capable on complex reasoning |
| Qwen3 8B | Coding, multilingual | Strong general-purpose pick, especially outside English |
| Phi-4-mini | Low-resource reasoning | Punches above its size on reasoning tasks per its size class |
12-16GB tier: mid-range GPU or a 16-24GB Mac
| Model | Best for | Notes |
|---|---|---|
| Qwen3 14B / Qwen 2.5 14B | Coding, analysis | Community favorite at this tier for code-adjacent work |
| DeepSeek-R1-Distill 14B | Step-by-step reasoning | Distilled from DeepSeek's reasoning model; strong at math and logic chains |
| Gemma 3 12B | Multimodal, general use | Handles image input alongside text; well-rounded |
| Phi-4 | Reasoning per gigabyte | Smaller footprint than most peers at similar reasoning quality |
24GB tier: RTX 4090/5090-class GPU
| Model | Best for | Notes |
|---|---|---|
| Qwen3 32B | Reasoning, coding | The clear step up from the 14B tier in capability |
| Gemma 3 27B | Multimodal, general use | Largest Gemma 3 size that comfortably fits this tier |
| DeepSeek-R1-Distill 32B | Advanced reasoning, math | The strongest local reasoning most people can actually run |
| Mistral Small 3.1 (24B) | Fast execution, translation | Apache 2.0 licensed, multimodal, 128K context |
Qwen shows up at every tier above, which is why it gets its own dedicated run guide covering version choice and two setup gotchas worth knowing before you download it.
Apple Silicon: what unified memory unlocks
Because unified memory is one shared pool, a Mac with enough total memory can run models that wouldn't fit any consumer discrete GPU at all. A 128GB Mac Studio (M5 Max) or higher opens up very large open-weight models, including mixture-of-experts models like Llama 4 Scout, which totals around 109B parameters but only activates a fraction of them per token. The catch: MoE models still need their full parameter set loaded into memory even though fewer parameters compute per token, so the memory requirement tracks the total size, not the active size. Check the specific model's total parameter count against your unified memory in the calculator before assuming an MoE model is "cheaper" to run than its active-parameter count suggests.
When bigger isn't better
A model that barely fits your hardware at Q4 with no headroom for context is often a worse experience than a smaller model with room to breathe. Long conversations and large documents need context memory on top of the model's base weight size (see the calculator's context setting), so if you regularly work with long inputs, it's usually better to drop one tier and keep the headroom than to max out capacity on the biggest model that technically loads. Even the biggest model that fits your hardware still won't match a cloud model on genuinely broad questions; see what local AI can't do well for where that gap actually shows up.