Larger models (7B+) can run but often cause slowdowns, memory pressure, and inconsistent performance. For the smoothest experience on mid-range CPUs + 16GB, stick to models under 3B parameters with Q4_K_M quantization.
The best overall choice for i5 + 16GB systems. Llama 3.2 1B delivers impressive quality while using just ~1.5GB RAM. Fast inference, excellent instruction following, and 128K context support.
Perfect for chat, summarization, quick tasks, and running multiple models simultaneously.
Qwen 2.5 1.5B punches way above its weight class. Excellent at reasoning, coding basics, and multilingual tasks. Uses only ~2GB RAM while delivering quality that rivals larger models.
Outstanding instruction following and structured output generation.
The coding-specialized variant of Qwen 2.5. Excellent at code completion, explanation, and debugging. Surprisingly capable for a 1.5B model — handles Python, JavaScript, and more.
Perfect for developers who want a fast, local coding assistant without heavy resource usage.
SmolLM2 is specifically designed for efficiency. At 1.7B parameters, it offers excellent performance for its size with strong instruction following and reasoning capabilities.
Great balance between capability and resource usage.
TinyLlama remains a solid choice for ultra-lightweight inference. Trained on 3 trillion tokens, it delivers good quality chat responses while using minimal resources (~1GB RAM).
Excellent for embedded systems, mobile, or running multiple models at once.
EXAONE 3.5 2.4B from LG AI Research offers impressive capabilities in a compact package. Strong reasoning and instruction following with excellent Korean and English support.
One of the most capable models under 3B parameters.
Microsoft's Phi-2 at 2.7B parameters delivers strong reasoning and coding capabilities. While slightly older, it remains competitive and runs smoothly on i5 systems.
Excellent for logical reasoning, math, and structured tasks.
StableLM 2 Zephyr is fine-tuned for conversational AI using Direct Preference Optimization (DPO). Delivers natural, engaging dialogue while using minimal resources.
Great for chat applications and interactive assistants.
| Model | Params | RAM (Q4) | Best For | Speed |
|---|---|---|---|---|
| Llama 3.2 1B | 1B | ~1.5GB | General, Chat | ⚡⚡⚡⚡ |
| Qwen 2.5 1.5B | 1.5B | ~2GB | Reasoning, Writing | ⚡⚡⚡ |
| Qwen 2.5 Coder 1.5B | 1.5B | ~2GB | Coding | ⚡⚡⚡ |
| SmolLM2 1.7B | 1.7B | ~2GB | Efficiency | ⚡⚡⚡ |
| TinyLlama 1.1B | 1.1B | ~1GB | Ultra-light | ⚡⚡⚡⚡ |
| EXAONE 3.5 2.4B | 2.4B | ~3GB | Complex Tasks | ⚡⚡ |
| Phi-2 | 2.7B | ~3GB | Reasoning, Math | ⚡⚡ |
| StableLM 2 Zephyr | 1.6B | ~2GB | Conversation | ⚡⚡⚡ |
For mid-range CPUs + 16GB systems, use Q4_K_M quantization:
All download links above include Q4_K_M variants. Look for files ending in Q4_K_M.gguf.
Best Overall: Llama 3.2 1B — fast, capable, minimal RAM
Best for Quality: Qwen 2.5 1.5B — punches above its weight
Best for Coding: Qwen 2.5 Coder 1.5B
Best for Speed: TinyLlama 1.1B — instant responses
All models run completely offline with full privacy. Your data never leaves your device.