Local LLM for Low-RAM Laptops

You don’t need 32 GB of RAM to run a useful local AI.
If your laptop has 8–16 GB RAM and no dedicated GPU, stick to models under 3B parameters—ideally 1–2B. Big models (7B+) will crawl or crash.

Here’s the shortlist for 2025:

---

✅ Gemma 3 (1B) Best for ultra-light setups

• Runs on 8 GB RAM (even less with 4-bit quantization)
• 128K context great for long docs
• Text-only, but fluent, fast, and open (Google, Apache 2.0)
  → Use if: You need basic Q&A, summarization, or coding help on an older machine.

✅ SmolLM2 (1.7B) Best quality in small size

• Needs ~12–16 GB RAM
• Beats most 1–2B models in reasoning and code
• Open weights, fine-tuned for on-device use
  → Use if: You want the smartest small model that still runs on CPU.

✅ DeepSeek R1 (1.5B) Best for math & logic

• Built for problem-solving (outperforms GPT-4o on AIME math)
• ~16 GB RAM recommended
• Distilled from Qwen—lean but sharp
  → Use if: You’re coding, debugging, or solving technical problems offline.

---

7B Models: Possible, but with Limits

• Llama 3.1 8B, GLM-4 9B, Qwen2.5-VL 7B
  → These might be heavy for low-RAM laptops. Save them for machines with more than 16 GB or a GPU.

Some quantized 7B models (like Mistral 7B, Llama-3.2 3B, or Phi-3-mini 3.8B) can run on 16 GB RAM using LM Studio or Ollama, if:

• You’re using 4-bit quantization (Q4 or lower)
• You close other apps (Chrome is a RAM vampire)
• You accept slower responses (1–3 words/sec on CPU)

→ Use 7B only if: You need slightly better reasoning and you’ve tested it on your exact setup.

→ Stick to ≤3B if: You want predictable performance during travel, low battery, or multitasking.

---

🛠️ How to run them:

1. Install Ollama or LM Studio (free, open, one-click setup)
2. Search for the model name (e.g., gemma3:1b, smollm2:1.7b-q4)
3. Run. No cloud. No API keys. No bills.

---

Local AI isn’t about power, it’s about freedom.
Pick small. Run offline. Own your workflow.

[Top]