The 2026 Model Library

Families that matter right now

Family	Maker	Sizes	Why care
Llama 3.x	Meta	1B–405B	Strong general-purpose; reliable tool use from 3.1+
Qwen 2.5 / 3 / 3.5	Alibaba	0.5B–235B	Multilingual; tools; Qwen 3 has hybrid thinking
Gemma 3	Google	270M–27B	Strong quality-per-size; vision — check Gemma 4 before settling
DeepSeek-R1	DeepSeek	1.5B–671B	Reasoning model; competitive with O3 on math
Phi-4	Microsoft	14B	State-of-the-art for its size
Mistral / Mixtral	Mistral	7B–12B / 8x7B	Tool support; Nemo has 128K context
GPT-OSS	OpenAI	20B–120B	Open-weight; tools; thinking
Qwen3-Coder	Alibaba	30B–480B	Coding / agentic specialization

How to pick

Start with the default Llama or Qwen family at the size your hardware fits. They're well-tested, well-supported, and have reliable tool-calling.
Move to a specialist when the general one fails. Coding-heavy work → Qwen3-Coder. Hard math / reasoning → DeepSeek-R1. Vision → Gemma 3 or Qwen 2.5-VL.
Don't chase newest. Stable models that have been on the library for a few months have community Modelfiles, GGUF variants from bartowski, and vetted prompt templates.

The library is a moving target

This list will be partly out of date by the time you read it — new families land monthly. The shape of the decision (general vs specialist; size vs quant; tool capability) is more durable than any specific model name.

EV v3 · 2026-09-18: Two things in this lesson never worked, both since publication. The capability filter piped ollama show through grep -i capabilities, which matches the header and throws away the rows under it — the loop printed the word “Capabilities” three times and nothing else. The same mistake sat in the tools track's capability check and in the vision track's model list, and all three now read the block with awk (or ask /api/show for .capabilities). This loop needed one more fix: it queried model names this lesson never pulls, and ollama show only reads the local store — it 404s on anything you have not pulled — so it now walks ollama list. Separately, the daily-driver pull said qwen3-coder:7b, a tag that has never existed: the family table two screens above already says 30B–480B, and the library lists only 30b and 480b. The coding specialist here is now qwen2.5-coder:7b, which is real, fits the same hardware, and — unlike Qwen3-Coder — carries a fill-in-the-middle template, which matters for the /api/generate lesson.

Code

Browse the library, then commit·bash

# Capability filter over models you ALREADY have. `ollama show` reads the
# local store and 404s on anything you have not pulled, so walk
# `ollama list` rather than a wish list.
#
# `grep -i capabilities` would match only the header row and drop the
# capabilities themselves, so read the block under it.
for m in $(ollama list | awk 'NR>1 {print $1}'); do
  printf '%-28s ' "$m"
  ollama show "$m" | awk '/^  Capabilities/{f=1;next} f&&NF==0{exit} f{print $1}' | paste -sd' ' -
done
# vision / tools / thinking / insert (fill-in-the-middle) are the ones
# worth filtering on.

# Same answer straight from the API, if you have jq:
curl -s http://localhost:11434/api/show -d '{"model":"qwen2.5:7b"}' \
  | jq -r '.capabilities[]'

# Pick ONE general model + ONE specialist as your daily set
ollama pull qwen2.5:7b           # general
ollama pull qwen2.5-coder:7b     # coding specialist (Qwen3-Coder starts at 30B)

Families that matter right now

How to pick

The library is a moving target

Code

External links

Exercise

Progress

Comments 0