Guides · Updated
Ollama models explained: picking a model by size and task
What names like qwen3.5:9b, gemma4:e4b and 30b-a3b mean, how model size maps to memory, and which kind of model to use for chat, coding, reasoning, images and documents.
The Ollama library has a lot of models, but picking one comes down to two questions: what fits in your memory, and what job you need done.
Reading a model name
Ollama models are named family:tag, for example qwen3.5:9b.
- Family (
qwen3.5,gemma4,llama3.2): the model series and generation. A higher version number is usually newer within a family, but don’t compare numbers across families. - Size (
4b,9b,27b): parameters in billions. Bigger models are smarter, need more memory and run slower. - No tag (
ollama run gemma4): Ollama uses thelatesttag, which points to the model’s default size (forgemma4that’se4b). Name the size explicitly so you know what you’re getting. - Two numbers (
30b-a3b, “35B-A3B”): a mixture-of-experts (MoE) model with 30B parameters in total, of which about 3B are active for each token. - Other labels:
e2b/e4bon Gemma 4 mean effective parameters (small models built for laptops and low-power devices);coder/codingmeans a coding tune;vlmeans vision-language. - Quantisation tags: the default tags are usually 4-bit. Many models also offer higher-precision versions such as
q8_0; see the Tags tab on the model’s page at ollama.com. Exact tag names vary by model.
Size: what fits where
Download sizes for the default 4-bit tags (from our GPU picker data, checked October 2026):
| Size class | Typical download | Runs fully on | Examples |
|---|---|---|---|
| 1-4B | 1-3.5 GB | Any laptop, CPU-only PCs, 4-6 GB GPUs | llama3.2:3b, qwen3.5:4b, phi4-mini |
| 7-9B | 5-7 GB | 8 GB GPUs, 16 GB Macs | qwen3:8b, qwen3.5:9b, ministral-3:8b |
| 12-14B | 8-9.5 GB | 12 GB GPUs | gemma4:12b, qwen3:14b, phi4 |
| 20-24B | 14-15 GB | 16 GB GPUs (tight) | gpt-oss:20b, mistral-small3.2, devstral-small-2 |
| 27-32B | 17-20 GB | 24 GB GPUs, 32 GB Macs | qwen3.6:27b, gemma4:31b, deepseek-r1:32b |
| 70B and up | 40 GB+ | 48 GB+ VRAM or 64-128 GB unified memory | llama3.3:70b, gpt-oss:120b |
Two rules of thumb:
- Leave headroom. Budget about 1.5 GB on top of the download size for a short conversation, and more for long documents.
- Go as big as fits. A bigger model at 4-bit generally beats a smaller one at 8-bit. The details are in VRAM and quantisation explained.
MoE models are the exception to “bigger is slower”. They need memory for all their parameters but run at roughly the speed of their active parameters. That’s why gpt-oss:20b, gemma4:26b and qwen3-coder:30b feel fast for their size, and why they still run reasonably when split between GPU and system RAM.
Task: which kind of model
| Job | Look for | Good examples |
|---|---|---|
| Everyday chat & writing | General all-rounders | qwen3.5, gemma4, ministral-3, mistral-small3.2 |
| Coding chat & edits | Coding tunes | qwen2.5-coder, qwen3-coder, devstral-small-2, qwen3.6:27b-coding |
| Tab autocomplete | Tiny, fast coder | qwen2.5-coder:1.5b |
| Maths, logic, planning | Reasoning (“thinking”) models | deepseek-r1, qwen3, phi4-reasoning, magistral, gpt-oss |
| Images & screenshots | Vision models | qwen3-vl, gemma4, ministral-3 |
| Chat with documents | Long context, plus an embedding model | qwen3.5 or ministral-3 with nomic-embed-text |
| Old or low-power machines | Small models | gemma3:1b, llama3.2:3b, qwen3:4b |
Vision models read images; they don’t create them. To generate pictures, see Generating images locally. On an older or low-power PC, see Local AI on a budget PC.
A few things that trip people up:
- Reasoning models think out loud. They write out their working before answering. That helps on maths and multi-step problems but makes simple chat slower. For quick questions, an all-rounder often feels better.
- Embedding models aren’t chat models.
nomic-embed-textandqwen3-embeddingturn text into vectors for search. You use them through the API or an app, as in our local RAG guide. - Agents need tool support. Coding agents and automations call tools, so pick a model with the tools capability. Ollama shows capabilities like tools, thinking and vision on each model’s library page. Before you let one run commands, read Running local AI agents safely.
More on coding setups: local coding assistant guide.
Check a model before you rely on it
ollama show qwen3.5:9b # parameters, context length, quantisation, capabilities
ollama show qwen3.5:9b --license # the model's licence
ollama list # what you've downloaded, and how big it is
ollama rm qwen3.5:9b # delete a model you no longer need
Licences differ between model families, and some come with usage terms. Read the licence on the model card before using a model commercially.
Updating models, freeing disk space and clearing out old ones: Keeping local models up to date and tidy.
Beyond the Ollama library
Ollama can also run GGUF models straight from Hugging Face with ollama run hf.co/{username}/{repository}. Prefer the original publisher’s repository or well-known uploaders, for the reasons in our privacy guide.
Want a model trained to your own format or style? Fine-tuning a small model with LoRA explains when that’s worth it.
The short version
- Find your memory: GPU VRAM, or unified memory on a Mac.
- Pick the job: chat, coding, reasoning, vision or documents.
- Take the biggest model for that job that fits with headroom.
The GPU picker does all three for you. On a laptop? Read running local AI on a laptop first.
Get the Local AI weekly
New models worth running, what fits which GPU, and setup fixes. One short email a week, free. Unsubscribe any time.
Local-LLM App Starter
A tested starter for building your own app on Ollama: chat, RAG and a GPU-fit checker. Skip the setup hours.
kits.lilaforge.com → VibeEngineBest local LLM tools, ranked
Ollama, LM Studio, Open WebUI and llama.cpp scored from public evidence, with top picks.
vibeengine.lilaforge.com →