TL;DR: The best Ollama models depend on the job: small 1B to 3B models for lightweight tasks, 7B to 14B for everyday chat and RAG, a 30B MoE like Qwen3-coder for coding, and DeepSeek-R1 distills for reasoning. This guide covers the model library, the best Ollama models by task, and how to size GPU and VRAM to run them reliably for a business.
Ollama has become shorthand for the growing catalog of open LLMs (Llama, Qwen, DeepSeek, Mistral, Gemma) pre-packaged to run on your own hardware. Because they run locally, nothing you send through them leaves your network. Here is which model to run, by task, and the hardware you need.
What are Ollama models?
Ollama is a runtime that packages open-weight models with their configuration, prompt template, and license into a single artifact. Instead of downloading raw weights and wiring up a tokenizer by hand, you pull a ready-to-run model with one command and Ollama handles quantization, GPU offload, and serving.
Pulling and running a model
The official library lists well over 200 models. Getting one running takes two commands: ollama pull qwen2.5:14b to download it, then ollama run qwen2.5:14b to load it and open a prompt. Behind the scenes, ollama run also exposes the model over a local API on port 11434. Use ollama list, ollama rm, and ollama show to manage what is installed.
The best Ollama models by task
“Best” depends on the job. Here are the strongest picks by task as of 2026:
General chat & assistants
Llama 3.1:8b or Qwen2.5:14b, well-rounded defaults that run on a single mid-range GPU.
Coding
Qwen3-coder:30b, a mixture-of-experts model with long context for whole-repo reasoning, fast for its size.
RAG
Qwen2.5:7b/14b as the generator, plus nomic-embed-text or mxbai-embed-large for embeddings.
Summarization
Qwen2.5:14b/32b for long documents; Mistral Small 3.1 for fast, high-volume jobs.
For reasoning-heavy work such as code review, risk analysis, and multi-step tasks, the DeepSeek-R1 distills (8b/14b/32b) add a chain-of-thought trace that helps catch details a plain model would skip.
Sizing hardware: GPU and VRAM by model size
The biggest driver of quality and cost is how many parameters fit in fast memory. At 4-bit quantization, budget roughly half a gigabyte of VRAM per billion parameters, plus context overhead:
- 7B to 8B: 8 to 12 GB VRAM; a single consumer GPU or Apple Silicon Mac.
- 13B to 14B: 16 GB for headroom with longer context.
- 30B MoE (Qwen3-coder:30b): 20 to 24 GB, since only a fraction of parameters activate per token.
- 32B dense: 24 GB and up; an RTX 4090/5090-class card.
- 70B: 48 GB, or two 24 GB GPUs.
For most business assistants and RAG workloads, a 14B to 32B model on a single well-specified GPU is the sweet spot of quality, latency, and cost.
Understanding quantization tags
The tag after the colon (e.g. q4_K_M) describes how compressed the weights are. Lower bits use less memory and run faster at a small quality cost:
- q4_K_M: the default sweet spot for most business deployments.
- q5_K_M / q6_K: more fidelity for nuance-sensitive tasks.
- q8_0 / fp16: near or full precision when memory is not the constraint.
Running Ollama models for your business
Everything above works the same on a laptop and a server. The difference between an experiment and a deployment is what surrounds the model: authentication on the API, GPU capacity for concurrent users, monitoring, backups, and safe updates. That is the gap a private Ollama deployment for business closes: picking the right model and quantization, sizing the GPU correctly, and hardening the API, all on infrastructure you control, as part of our private and on-premise AI practice.
Want the right Ollama models chosen and run privately for your team?
Book an AI strategy session →A safe default for most teams
For most businesses, a 14B to 32B open model on a single well-specified GPU covers chat, RAG, coding, and summarization. Run it privately with Ollama so your prompts and documents never leave your network.