The most important step in the local-AI journey. At 8B, tool calling becomes reliable — this is the first model that genuinely remembers you across chats, reads your files, and browses the web. It stops being a chat toy and becomes an assistant that does things.
New here? The two-minute concepts primer covers the jargon — it applies to every model, so it isn't repeated here.
Where hardware starts to matter. Because it can use tools, conversations grow longer (file contents and tool results all consume context), so a roomy context window helps.
Raise the context window in .env — especially important here, since tools fill it fast:
Everything that failed on the chat-only models should start working here. Tool calling is ~85% reliable — when a call misses, naming the tool gets it back on track.
The single biggest capability jump in the whole lineup. This should just work.
recall and accurately reports autumn + Hirakata in the new chat. The memory persists because it lives in Aperio's database, not the conversation.Llama 3.1's training stops around December 2023 — older than the Qwen models. But it can fetch the live web, which more than makes up for it.
fetch_url and returns the current answer with a source — instead of stopping at its 2023 knowledge.The day-to-day reason you stepped up to 16 GB: the model storing context, writing documentation, and reading your project — in one go.
remember, then wiki_write (it may ask you to approve the write), then scan_project / read_file. It should attempt all three and succeed at at least two.A generalist — competent at multi-step reasoning, if not as deep as the larger tiers.
Llama 3.1 8B is the threshold model: the smallest one that turns Aperio from a chat window into an assistant. It's a reliable generalist that trades a little creative flair for solid tool use. If 16 GB is your ceiling, this is the best experience within it. Creative writing your priority instead? A Qwen 4B may actually please you more.