Minimax H3 Is Having a Moment. My Prompt Set Doesn't Care.

This morning my feed looked like everyone had the same assignment. Minimax H3 every other post. Cheaper. Better. More open. I won't pretend the hype didn't hit. It always hits. I started opening tabs.

Then I caught myself. I know that pattern too well. A new release shows up, the threads say cheaper and better, and I start adjusting prompts that worked before the hype arrived. So before I argue about Minimax H3, I run it through the one thing I trust: my own tiny prompt set.

My prompt set is boring and personal

It's not a benchmark. It's eight cases I keep around because they resemble real work:

• repair malformed JSON without dropping keys
• turn a rambling function into plain logic
• write SQL for duplicate emails
• summarize an error log in three lines
• refuse a bad request without a lecture
• call a tool instead of explaining it
• keep the middle of a long instruction
• stay under a loose latency budget

Boring? Yes. That's the point. A model can win a leaderboard and still mangle my JSON repair case. I don't need it to be brilliant. I need it to not ruin my Tuesday.

The runner is provider-agnostic on purpose

I wrote a small runner where the provider is just a function. I don't want my test to depend on any vendor's SDK or dashboard. That way the value is in the test set, not in the integration.

Here's the skeleton I've been using. I've removed my scoring helpers so it stays readable:

from dataclasses import dataclass

@dataclass
class SmokeCase:
    name: str
    prompt: str
    signal: str

SMOKE_SET = [
    SmokeCase(
        'json_repair',
        "Fix this and return only valid JSON: {'a': 1, 'b': [2, 3,}",
        'Output starts with { and parses as JSON',
    ),
    SmokeCase(
        'tool_call',
        "Use get_weather(city). What's the weather in Lisbon?",
        'Response is a tool call, not a paragraph',
    ),
    SmokeCase(
        'long_middle',
        'List three rules, then output the middle rule after this long filler: ignore everything before the marker. The middle rule is: never log plaintext passwords.',
        'Mentions passwords, not the filler',
    ),
]

def run_smoke(case, provider):
    reply = provider(case.prompt)
    return {'case': case.name, 'reply': reply, 'signal': case.signal}

I run this with temperature 0 and the exact same prompt every time. If I change one word, I've broken the comparison. This is the part people skip when they're excited. It's also the part that keeps me honest.

Free access is what makes this sustainable

Here's where MonkeyCode's free model access and free server option become useful to me.

Disclosure: This article was prepared as part of MonkeyCode's product outreach.

I'm not using free access to try every model on the internet. I'm using it to run the same boring smoke set without deciding whether each release is worth my API budget. That matters because I used to skip the negative cases when testing cost money. I'd run the happy path and tell myself the rest was probably fine. The refusals, the malformed JSON, the long prompts — those are exactly where models surprise me.

The free server option makes the workflow easier to share. I can put this tiny runner somewhere without spinning up a paid box, point it at a provider, and get a yes/no answer. Same code, same cases, different model. That feels closer to how open-source tools are supposed to work: small, replaceable, and easy to run yourself.

"Open-source spirit" means replaceability, not a sticker

I keep hearing open used as a vibe, so let me say what I mean by it.

For a model release, I care about the part that changes my daily workflow:

• Can I run it through a code path I control?
• Can I swap it out without rewriting the runner?
• Can I see where it fails instead of guessing from a table?
• Can another developer reproduce my test without buying something first?

That's it. Not a license argument. Not a philosophy lecture. Just the practical ability to replace a component and observe what happens.

When a thread only tells me the model is cheaper and better, I have nothing to reproduce. When a free option lets me run my own cases and watch it fail, I can make an actual decision.

What I won't claim

This smoke set won't predict your production traffic. Eight cases is eight cases, not a benchmark. Free access doesn't mean unlimited, and I don't have visibility into quotas, SLAs, or long-term availability. I'm not saying any specific model passed or failed. One quiet afternoon run is not science.

Skip this approach if you need hard latency guarantees, compliance approval, or production support. A free tier is not a contract. If you need a global ranking, this won't give you one. It gives you something smaller but more useful: a repeatable signal about whether a model fits your work.

My takeaway

The Minimax H3 moment will pass. Next week there will be another one. My prompt set won't care which release is trending. It will just keep catching the same practical failures — and that's exactly why I keep running it.

If you maintain a tiny model smoke set, what's in it? I especially like the boring, personal cases that never make it into benchmarks.