Local LLM model picker: how to choose between Llama, Mistral, Qwen and DeepSeek

anant94 · reddit · 2026-09-13

A practical decision framework for choosing which open-source model to run locally with Ollama or LM Studio, arguing that benchmarks rarely reflect day-to-day behavior on your hardware.

Llama (Meta): reliable all-rounder. 8B fits most laptops for chat/RAG; 70B is the sweet spot for complex reasoning if you have VRAM. Safe commercial licensing.
Mistral: speed and parameter efficiency; 7B/12B models excel at low-latency, high-throughput pipelines and summaries.
Qwen (Alibaba): underrated for multilingual workloads, structured JSON output, function calling, and heavy math/reasoning (14B–72B).
DeepSeek: coding specialists — its 6.7B/14B coder models often beat generic 13B/70B models at code generation and review.

Decision rule: coding → DeepSeek-Coder/Qwen-Coder; general chat/RAG on modest hardware → Llama-3-8B or Mistral-7B; multilingual or strict structured data → Qwen-2.5-14B/32B; complex multi-step logic → 70B-class models.

VRAM rules of thumb (4-bit quant): 7–8B ≈ 5–6 GB (M-series MacBook 16GB or 8GB GPU); 14–32B ≈ 10–20 GB (RTX 3090/4090 or 32GB+ Mac); 70B ≈ 38–42 GB (dual GPUs or 64GB+ unified memory).

Part 1 of a 3-part series; later parts cover quantization benchmarks and production deployment.

Original post →

Local LLM model picker: how to choose between Llama, Mistral, Qwen and DeepSeek

More from Infra