Confirmed working. No account needed.
This page walks through running OpenCode and picking a free model, such as Big Pickle or MiMo-V2.5. It assumes OpenCode is already installed on the machine.
Why this page exists:
It’s a departure from the rest of the track, on purpose. The one standing rule it breaks: one command. The track uses
claude; this page usesopencodeinstead. It needs no account and no API key, so it’s a smaller departure than it looks. That’s fine for a take-home page. It’s not fine inside the Session 1 spine — don’t fold it in.
Start OpenCode in a project folder, the same way campers start claude:
cd $HOME\Documents\Projects
cd <Name>
mkdir opencode-test
cd opencode-test
opencode
That’s the whole setup.
/connect.Inside OpenCode:
/models
Pick one of the free models. As of writing, the free list is:
| Model | ID |
|---|---|
| Big Pickle | big-pickle |
| DeepSeek V4 Flash | deepseek-v4-flash-free |
| MiMo-V2.5 | mimo-v2.5-free |
| Nemotron 3 Ultra | nemotron-3-ultra-free |
| North Mini Code | north-mini-code-free |
| Laguna S 2.1 | laguna-s-2.1-free |
| Ling 3.0 Tiny | ling-3.0-tiny-free |
| LongCat 2.0 | longcat-2.0-free |
Writing out a model name in full: opencode/ plus the ID. For example,
opencode/big-pickle or opencode/mimo-v2.5-free.
Note the pattern: every free model except Big Pickle carries -free in its ID.
That suffix is the difference between free and billed — check it twice.
Don’t assume the connection is good just because nothing errored. Run the track’s own standard, and make it build something you can check:
Build a dice roller in Python with a tkinter window. Big button that says ROLL, and
when I click it, it shows a random number from 1 to 20 in huge text. Save it as
dice.py
Then:
python dice.py
claude and through a free Zen
model, then look at the two results side by side.| What you see | Likely cause |
|---|---|
| Works, then stops mid-class | Rate limit. Switch models or wait it out |
Model missing from /models |
Free trial for that model ended — pick another from the list |
| Nothing connects, on every machine | Lab firewall blocking outbound access — nothing you can fix in the room |
The general troubleshooting reference is
troubleshooting.html. It’s written for claude, but
the sections on hung sessions and on reading errors apply to any agent in a terminal.
These models were free as of August 10, 2026.
| Rank | Model | Parameters (Total / Active) | Context Length | Key Coding Benchmarks |
|---|---|---|---|---|
| 1 | DeepSeek V4 Flash | 284B / 13B | 1M tokens | Terminal Bench 2.1 82.7%, DeepSWE 54.4%, Toolathlon Verified 70.3%, NL2Repo 54.2% [github.com], [docs.zenmux.ai] |
| 2 | Nemotron 3 Ultra | 550B / 55B | 1M tokens | Code benchmark 85.3%; designed for long-horizon agentic coding, reasoning, and software engineering tasks. [open.bigmodel.cn], [github.com], [freeaiapi.org] |
| 3 | LongCat-2.0 | 1.6T / ~48B active | 1M tokens | SWE-Bench Pro 59.5%, Terminal-Bench 70.8%. Purpose-built for agentic coding. [en.oninvest.com], [en.webhakim.com] |
| 4 | Laguna S 2.1 | 118B / 8B | 1,048,576 tokens | SWE-Bench Multilingual 78.5%, SWE-Bench Pro 59.4%, Terminal-Bench 2.1 70.2%. [mastra.ai], [huntscreens.com] |
| 5 | Big Pickle (identified as GLM-4.6) | ~355B / ~32B active | 200K tokens | OpenCode maintainers confirmed Big Pickle is GLM-4.6; widely regarded as a strong coding-agent model. 200K context and 128K output. [huggingface.co], [atlas.kevinhu.io], [artificial…nalysis.ai] |
| 6 | MiMo-V2.5 | 310B / 15B | 1M tokens | Coding Agent score 71.8, MiMo Coding Bench 62.3, Terminal-Bench 2.0 56.1. [anyrouter.dev], [beatriz.page], [github.com] |
| 7 | North Mini Code | 30B / 3B | 256K tokens | Artificial Analysis Coding Index 33.4-41.7; built specifically for software engineering workflows. [cryptopolitan.com], [edgen.tech], [opencode.ai] |
| 8 | Ling-3.0-tiny | 7.9B / 1.3B | 256K-262K tokens | Artificial Analysis Coding Index 26.5, SciCode 24.2%. [brunch.co.kr], [hunted.space] |
This Vibe Coding Curriculum by Brian Bird, created in with AI assistance, is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.