12 Comparing and selecting LLM models

This chapter offers a practical way to navigate the crowded Ollama model library by focusing on the major open-weight families, their strengths, and the real-world tradeoffs that matter on a Mac. It introduces Llama, Gemma, Qwen, Mistral, and DeepSeek as the main options you are likely to encounter, then emphasizes that model names and rankings change quickly, so it is important to check what is installed locally and verify current tags before choosing a model.

The central decision is size versus quality: larger models generally perform better, but they also run slower and require more memory. The chapter explains how parameter counts, token throughput, and RAM limits shape what is practical on different machines, showing that 2B–3B models are often the sweet spot for 8 GB Macs, while 7B–8B models become reasonable on 16 GB systems. It also shows why benchmarks are useful but incomplete, and why the best choice is ultimately the model that works well for your specific tasks and hardware.

Language support is another key selection factor. Qwen is presented as the strongest choice for Asian languages, especially Japanese, Mistral as the best fit for European languages, and Llama as a reliable all-rounder for English and general use. The chapter then shows how easy it is to switch models in a local chatbot by changing a single model name or adding a sidebar selector in Streamlit, and it closes with a look at February 2026, when new mixture-of-experts models dramatically improved the open-model landscape by delivering frontier-level capability while keeping active computation relatively low.

The five major open-weight model families discussed in this chapter and how they connect to your local machine through Ollama

Multilingual capability comparison across model families. Qwen excels at Asian languages, Mistral at European languages, and Llama is a strong all-rounder.

The February 2026 benchmark landscape, showing open models reaching parity with proprietary models on the hardest benchmarks. Reported results use different evaluation setups and should not be treated as a controlled head-to-head comparison.

• Five major model families dominate the open-weight LLM landscape: Llama (Meta), Gemma (Google), Qwen (Alibaba), Mistral (Mistral AI), and DeepSeek (DeepSeek AI). Each has distinct strengths.
• The model size (number of parameters) directly affects quality, speed, and RAM requirements. Bigger models are smarter but slower.
• All five families perform well in English, although results vary by model size and task. Performance in other languages differs more noticeably: Qwen provides particularly strong support for several Asian languages, Mistral emphasizes European-­language coverage, and Llama offers broadly capable multilingual performance.
• Switching models is trivial: change one line of code or add a UI dropdown. The Ollama API is the same for every model.
• Benchmarks are guides, not gospel: always test models with your own use case.
• February 2026 changed everything. Mixture of experts (MoE) models (GLM-5, Qwen3-Coder-Next, and MiniMax M2.5) achieved parity with proprietary models at a fraction of the cost, using far fewer active parameters than their total size suggests.
• MoE is the new paradigm. The total number of parameters represents knowledge capacity, but the number of active parameters determines the compute cost. The gap between them is the key to running frontier AI locally.

FAQ