← Free AI Model Comparison

Which free model should I use?

A one-glance decision chart for the free models on OpenCode Zen

LightDark

The 10-second answer: use DeepSeek V4 Flash for almost everything. Switch away only when you hit one of the five situations below.

Think of it like gears on a bike — you stay in one gear most of the time and change only when the road does.

Which free model? 1 Is it private or confidential? Client data, passwords, anything you'd not post online. Yes No STOP — don't use any of these Free = your data may train the model. Use local Ollama, or Claude. 2 Does the job involve a picture, screenshot, a video or audio? Yes No MiMo V2.5 The only one that sees and hears. Text, images, video, audio. 1M context. 3 Is it genuinely hard? Tricky maths, science, or a long job it has to plan out on its own. Yes No Nemotron 3 Ultra The heavyweight. 550B params, 1M context. Slowest, but it gets there. 4 Still thinking work — but you want it snappier? Good reasoning, less waiting around. Yes No Hy3 Punches above its weight on reasoning. 256K context — the small one here. 5 Do you want it running on your own PC? Offline, nothing leaves the machine. Yes No North Mini Code The only one you could self-host — but it wants a ~18GB graphics card. Smaller PC? Use Ollama instead. ✓ Anything else — everyday coding and chat. This is where you'll be 80% of the time. DeepSeek V4 Flash Your default driver. Fast, brilliant at code, reads 1M tokens at once. And "Big Pickle"? A mystery box — today it quietly hands your question to DeepSeek V4 Flash. Nothing to gain. Skip it.

When to change gear mid-job

You've started. Something's not right. Here's the switch to make.

Shift up
It's going round in circles, or keeps getting it wrong.
Move to Nemotron 3 Ultra. It's slower but it actually thinks the problem through. If that also stalls, the job probably needs Claude.
Shift down
It's slow and the job is simple.
Move to North Mini Code. Small, quick, made for straightforward code and terminal work.
Bigger tank
It's forgetting things, or your project is too big for it.
You've run out of context. Only DeepSeek V4 Flash, Nemotron 3 Ultra and MiMo V2.5 hold a million tokens. Hy3 holds a quarter of that.
Add eyes
You need to show it a screenshot, photo, video or a recording.
Only MiMo V2.5 can look and listen. The others read text only.
Middle gear
Needs proper reasoning, but Nemotron is overkill.
Hy3 is the sweet spot for maths and science without the wait.
Pull over
The work involves real client or personal data.
Stop using the free models. Run Ollama on your own machine, or use Claude. Free access is paid for with your data.
Why "free" comes with a catch: during the promo period, what you send to these models may be used to train them. Fine for your own side projects and experiments. Not fine for anything you wouldn't be happy posting in public.
None of these run on your computer. All six are large language models — a type of AI — running on someone else's servers; you reach them over the internet through OpenCode. Most are "open weights", meaning the file is published free, so in theory you could download and run one yourself. In practice only North Mini Code is small enough to try, and it still wants a serious graphics card (~18 GB). For AI that genuinely runs on a normal PC — offline, private, free forever — you want Ollama and much smaller models: here's how we set ours up.
Claude Project Dashboard · July 2026 · full model comparison · run one on your own PC