Why Volvox.Bot Runs on GLM (and How We Picked a Model)

Every Volvox.Bot feature runs through a language model. Here's how we tested our options, why we landed on GLM from Z.ai, and how to run the same kind of bake-off for your own project.

Eleftheria Batsou

Eleftheria Batsou

Developer Advocate

2026-11-17
15 views
6 min read
ProductAILLM
Why Volvox.Bot Runs on GLM (and How We Picked a Model)

One bot, a lot of jobs

Volvox.Bot does more than chat. It moderates, summarizes long threads, remembers people, and answers questions in busy servers at all hours. Every one of those features runs through a language model.

That makes the model choice one of the biggest decisions we make. It shapes how fast the bot replies, how smart it feels, and how much it costs to run for every server that adds it.

In the last post we promised to go deeper on this. So here's how we picked a model, why we landed on GLM from Z.ai, and how you can run the same kind of test for your own project.

What a Discord bot actually asks of a model

There's no single "best model." There's the best model for the job you're giving it. So before testing anything, we wrote down what each feature needs.

FeatureModel requirementsWhy it matters
AI chatSpeed, personality, staying on topicPeople expect a reply in seconds, in a busy channel
Auto-moderationAccuracy on context and sarcasmA false ban costs more trust than a missed spam message
TL;DR summariesLong context, getting the facts rightHundreds of messages in, a short, faithful summary out
User memoryFollowing instructions on what to keepIt should remember preferences, not private details
Tool use (tickets, commands)Reliable structured outputThe bot has to call the right action with the right fields
EverythingCost per messageEvery server that adds the bot adds to the bill

That last row is the one small teams forget. A model that's slightly smarter but five times the price can quietly decide whether a free tier is possible at all.

How we tested: a small, honest bake-off

Leaderboards and benchmark charts are a starting point, not an answer. They measure general skill. We needed to know how a model behaves in a Discord server at 2am when someone posts a wall of memes and one real question.

So the test looked like this:

  1. Build a test set from real situations. Typical chat questions, messy threads to summarize, borderline moderation cases, and messages that should not be flagged. Anonymized, of course.
  2. Run every candidate on the same set. Same prompts, same settings, no special tuning for one model.
  3. Judge blind where possible. Hide which model wrote which answer. It's surprisingly easy to favor a brand name.
  4. Measure what users feel. Time to first reply, not just total speed. A bot that starts typing fast feels faster.
  5. Do the cost math at scale. Estimate cost per 1,000 messages, then multiply by realistic server activity.

Why we landed on GLM

GLM is the model family from Z.ai, the international brand of Beijing-based Zhipu AI. It doesn't get as much attention as the big US labs. For a bot like ours, a few things made it hard to ignore.

  • Price for performance. Running a bot across many servers means millions of tokens. GLM gave us strong answers at a price that keeps a free tier realistic.
  • Long context. The latest GLM models handle a 1 million token context window (Z.ai docs(opens in new tab)). That's more than enough room to summarize a very long thread in one pass.
  • Tool calling and structured output. GLM supports function calling and JSON-schema responses (OpenRouter(opens in new tab)). That matters for tickets, commands, and anything where the bot has to take an action, not just talk.
  • Open weights. GLM-5.2 is released under the MIT license with weights on Hugging Face (model card(opens in new tab)). Even if we never self-host, it's good to know we're not locked in.

None of this means GLM wins every test. It means it was the best fit for our mix of features, budget, and scale.

The trade-offs we accepted

Every model choice comes with downsides. Here are the ones worth knowing about.

  • Reasoning costs time. Newer "thinking" models plan before they answer. Great for hard tasks, slower for a quick "hey bot, what's the server rules link?" Matching the model and effort level to the feature matters more than picking one model for everything.
  • Less community knowledge. There are fewer tutorials and forum answers for GLM than for the most popular models. Sometimes you're the one figuring it out.
  • Moving fast means re-testing. New GLM versions ship every few months. Each one is a chance to improve, and a chance for behavior to change. Keep your test set so you can re-run it.
  • Data and privacy questions. Any provider you send messages to needs to fit your privacy policy and your users' expectations. Check where data goes and how long it's kept, whatever model you pick.

Run your own bake-off

Building something with AI and not sure which model to pick? Here's the short version of our process:

  1. List your features and what each one needs. Speed, accuracy, context length, tool use, cost.
  2. Collect 50 to 200 real examples. Include the weird ones. Edge cases tell you more than easy wins.
  3. Shortlist three or four models. Mix at least one big-name model with one or two cheaper or open-weight options.
  4. Run them all the same way. Same prompts, same settings, same examples.
  5. Score blind. Have someone who doesn't know which model is which pick the better answer.
  6. Do the cost math at your real scale. Not per message. Per month, at the usage you expect.
  7. Keep the test set. You'll want it again the next time a new model comes out.

Next in Built in the Open

The best model on paper isn't always the best model for your product. Test with your own data, do the cost math early, and be ready to test again.

Next up: the mistakes we've made building with AI, so you don't have to make them too.

Want to see GLM in action? Add Volvox.Bot to your server(opens in new tab) and try the AI chat or TL;DR summaries yourself. Questions about how we tested? Come ask in our Discord(opens in new tab). No gatekeeping, no ego, just code.

Sources

Got a question about how we build? Want to share how your team uses AI? Join us on Discord(opens in new tab). No gatekeeping, no ego, just code.

More from the blog

Cookie Policy

We use cookies to improve your experience. Learn more