What Ollama Actually Is (and Isn't) | Edition 332

You're paying per message for Claude or ChatGPT — and for a lot of what you do, that's the right call. But not everything needs a frontier model. Formatting a spreadsheet. Summarizing a doc. Running a background automation at 2am. Does that really need to go to the same model that handles your hardest problems?

That's the gap Ollama fills. Here's what it actually is — and isn't.

What Ollama is

Ollama is a free, open-source way to run AI language models — either on your own computer, or, as of its newer "Ollama Cloud" option, on Ollama's own servers instead of yours.

Think of it like a media player, but for AI models instead of movies. Ollama isn't an AI itself — no personality, no opinions. It's the plumbing that loads a model and lets you talk to it. The models — Llama, Qwen, DeepSeek, Gemma, Hermes, and dozens more — are built by other labs (Meta, Alibaba, DeepSeek, Google, Nous Research) and released "open weight," meaning anyone can download and run them instead of only reaching them through one company's paid API.

It also plugs into tools you probably already use: Claude Code, Codex, OpenCode, VS Code, n8n. Point any of them at Ollama instead of — or alongside — a frontier provider, and your workflow doesn't change. Same commands, different (and usually much cheaper) model underneath.

What Ollama is NOT

• Not a chatbot brand. There's no "Ollama the assistant" the way there's "Claude the assistant." It's infrastructure for running someone else's model.
• Not automatically private just because "local" is its reputation. Run it on your own machine and nothing leaves it. Use Ollama Cloud instead, and your prompts go to Ollama's servers — they say they don't train on your data, but that's a policy promise, not a local guarantee.
• Not as strong as the best frontier models on the hardest problems. Open models have closed a lot of ground, but on genuinely hard reasoning or high-stakes coding, Claude and GPT-5-class models still generally lead.
• Not free at scale in the cloud. Local is free after setup. Ollama Cloud starts at $20/month (or $200/year), up to $100/month for heavier use — still far cheaper than frontier pricing for the same volume, but a real bill.

Where it sits in the landscape

• Frontier cloud AI (Claude, ChatGPT): best quality, closed models, you pay per use, your prompts go to that company's servers.
• OpenRouter: a marketplace, not a model. One account, 500+ models across 80+ providers — open and closed — at the same rate the underlying provider charges. No markup on usage.
• Ollama, local: free after setup, fully private, works offline — limited by your own hardware.
• Ollama, cloud: open models running on Ollama's hardware instead of yours. No GPU required, still far cheaper than frontier APIs, same setup as local.

OpenRouter is the toll road to hundreds of providers' models. Ollama is the tool that lets you run open models yourself — on your machine, or now, on theirs.

A few real models

Ollama ranks models by downloads. Near the top right now:

• Llama 3.1 (Meta) — still Ollama's most-downloaded model by a wide margin, the default a lot of people started with.
• Qwen 3.6 (Alibaba) — strong general-purpose and coding model, fast-growing.
• GLM 5.1 (Z.AI, formerly Zhipu) — built for coding agents and tool use.
• DeepSeek V4 Pro — strong reasoning per dollar, with large-context versions for big codebases.
• Hermes 4.3 (Nous Research) — tuned to follow your system prompt closely, deliberately without heavy built-in restrictions. Nous calls it "neutral alignment": guardrails are your choice, not baked into the weights. That cuts both ways — fewer refusals means it's a model you steer carefully, not one you hand to just anyone.

The names will change. The point is that "open model" means real variety, not one option.

When you'd actually use this

• Budget coding. Point Claude Code or Codex at an Ollama model for boilerplate and routine work; save frontier calls for the hard problems.
• Privacy-sensitive work. Legal docs, health records, internal-only data — run it fully local and nothing leaves the machine.
• Always-on automation. Background agents racking up a metered bill get dramatically cheaper on an open model. This is where dedicated agent apps come in too: Hermes Agent, Nous Research's open-source assistant with persistent memory (Discord, Telegram, WhatsApp, Slack) — a separate product from the Hermes model — and OpenClaw, an independent open-source personal-assistant project (not an Anthropic product, though built on Anthropic's Claude Agent SDK). Both are built to run on an open model instead of a metered one.
• Offline work. No signal on a flight, or a facility that blocks outside connections — local Ollama still works.
• Just experimenting. Free to install, free to run locally, no surprise bill while you learn.

The honest tradeoffs

Local needs a decent computer — RAM, ideally a GPU. Cloud is still a subscription that sends your prompts to a server, just not a frontier AI company's. Open models are catching up fast but still generally trail the best closed models on the hardest work. And it's one more decision to manage — "just use Claude for everything" is simpler than routing tasks by model.

Try this yourself

Pick one recurring task that doesn't need your best model — drafting a first pass, formatting a file, running a routine check — and try it on Ollama instead.

Download Ollama → It installs in under two minutes, and the free tier costs nothing to try.

If you already use Claude Code or another coding agent, point one session at an open model like Qwen or GLM for your next low-stakes task. Keep Claude for the problems that actually need it.

That's AI Super Simplified.