Skip to content
Rafael Minuesa edited this page Sep 14, 2026 · 3 revisions

AI-2 keeps a curated catalog of local models it knows how to run well. Everything is managed with the ai-2 model commands, see Commands.

The catalog

Model Parameters Download Peak RAM in use License
Gemma 3 270M Instruct 0.27B 242 MB ~650 MB Gemma Terms of Use
Qwen2.5 0.5B Instruct 0.5B 380 MB ~900 MB Apache 2.0
Qwen3.5 0.8B (MTP) 0.8B 525 MB ~1.2 GB Apache 2.0
Gemma 3 1B Instruct 1.0B 770 MB ~1.5 GB Gemma Terms of Use
Qwen3 1.7B 1.7B 1.06 GB ~2 GB Apache 2.0
SmolLM3 3B 3.0B 1.83 GB ~3.3 GB Apache 2.0
Qwen2.5 7B Instruct 7.0B 4.47 GB ~6.8 GB Apache 2.0

All models are Q4_K_M GGUF quantizations, downloadable anonymously (no Hugging Face account, no gated license click). The live catalog with checksums is ai2/data/models.yml; it updates with the ai-2 package through sudo pacman -Syu.

Embedding models (for your documents, not for chat)

ai-2 doc indexes your documents with a text embedding model, a different kind of catalog entry that is never offered for chat. Chosen 2026-09-14 on a Spanish test document (the Constitution, 37 parts, 12 questions, was the right part among the top 3):

Model Parameters Download Peak RAM while indexing Spanish test Used on
Nomic Embed Text v2 MoE (multilingual) 0.475B (0.305B active) 345 MB ~1.1 GB 12/12 Standard and up
Nomic Embed Text v1.5 (English) 0.137B 85 MB ~300 MB 11/12 Tiny and Light

Both Apache 2.0, first-party repositories, anonymous downloads. all-MiniLM-L6-v2 (the embedder the tier files used to name) found 4 of 12 and was not cataloged. Speed is the cost: on the 2016 reference laptop (2 cores, SSE4.2 build) either nomic model embeds about 9 tokens a second, so indexing is a "start it and come back" job there; see Commands.

Downloads you can trust

ai-2 model pull resumes interrupted downloads instead of restarting, verifies every file against the catalog's SHA-256 before accepting it, and refuses to start on a full disk. ai-2 model verify re-checks everything on disk any time.

How the recommendation works

ai-2 recommend combines two facts about your machine:

  1. What fits. Models whose measured peak RAM (loading spike and KV cache included, not just file size) fits your RAM with headroom left for the system.
  2. What runs. Each fitting model's speed is estimated from your measured benchmark, and the recommendation is the largest model that still runs at a usable speed.

If a larger model fits in RAM but would be too slow, or nothing clears the speed floor, AI-2 recommends remote inference for the heavy work instead of pretending, and says why. Local where possible, remote by explicit choice.

The recommendation never blocks you: if you decline its download, ai-2 chat and ai-2 serve simply use the best model already on disk.

The starter model

The ISO bundles Gemma 3 270M, so a fresh install chats with no network. AI-2 is plain about what it is, sub-1B models are fluent and fast but unreliable on facts and arithmetic. Once online, the wizard offers a model that genuinely fits your machine.

One measured caveat from the catalog: Qwen3.5 0.8B carries a multi-token-prediction head, but speculative decoding halves its speed on a compute-bound old CPU, so AI-2 does not enable it there.

Clone this wiki locally