Open-weight LLM models skills-set

llm_use_cases.md

continue.dev

Code Generation:

• qwen2.5-coder:7b - Excellent for code completion
• codellama:13b - Strong general coding support
• deepseek-coder:6.7b - Fast and efficient

Chat & Reasoning:

• llama3.1:8b - Latest Llama with tool support
• mistral:7b - Fast and versatile
• deepseek-r1:32b - Advanced reasoning capabilities

Autocomplete:

• qwen2.5-coder:1.5b - Lightweight and fast
• starcoder2:3b - Optimized for code completion

Qwen3

• math, coding, and reasoning (in mid-range hardware like Macbook Air M4)
• Specific model:
  • Qwen2.5-7B-Instruct: cited here by LM Studio team as perform well in a wide variety of tool use cases)

Gemma 3

• multilingual support
• factual knowledge
• vision capabilities

LLaMa 3.2

• basic tool-calling

Misc

Understanding Quantization

Quantization reduces the precision of the model's weights to save memory without significantly affecting its performance.

• FP16 (No Quantization): best
• INT8 Quantization: good
• INT4 Quantization: usable

How to Choose

• If you have a high-end GPU with plenty of VRAM, try FP16 first.
• For the best balance of quality and performance, use INT8.
• If you're on a system with limited resources, INT4 can make large models usable.
• You can always experiment with different quantization levels to find what works best

Sources:

• https://www.reddit.com/r/LocalLLaMA/comments/1kau30f/qwen3_vs_gemma_3/
• https://simonwillison.net/2025/Aug/10/qwen3-4b/
• https://spin.atomicobject.com/running-local-llms/
• https://www.caniusellm.com/
• https://docs.continue.dev/guides/ollama-guide#what-are-the-best-practices-for-ollama