Three cobblers with their wits combined surpass Zhuge Liang himself—Hermes MoA perfectly embodies this old saying. There's an ancient Chinese proverb: "Three cobblers with their wits combined surpass Zhuge Liang himself." Turns out, the wisdom of our ancestors was truly ahead of its time. And surprisingly, this saying still holds up in today's AI world. Case in point: Hermes has rolled out an MoA feature! Isn't this the perfect embodiment of "three cobblers with their wits combined surpass Zhuge Liang"? Can't get your hands on GPT-5.6? Can't use Fable5?

💡 What You Will Learn

Three cobblers with their wits combined surpass Zhuge Liang himself—Hermes MoA perfectly embodies this old saying. There's an ancient Chinese proverb: "Three cobblers with their wits combined surpass

Three Cobblers Beat Zhuge Liang — Hermes MoA Perfectly Embodies This Old Saying

There's an old Chinese saying: "Three cobblers with their wits combined equal Zhuge Liang, the mastermind." Turns out, the wisdom of our ancestors is truly ahead of its time. And surprisingly, this ancient proverb still holds up in today's AI world. Case in point: Hermes just rolled out a MoA feature! Isn't this the perfect embodiment of "three cobblers beat Zhuge Liang"? Can't get your hands on GPT-5.6? Or Fable 5? Give Hermes' newly launched MoA hybrid agent mode a shot!

GPT-5.5 is great at writing copy, but when I ask it to code, it gives me flashy but impractical stuff. DeepSeek V4 has strong reasoning, but its everyday conversational tone just feels a bit off. Having to manually switch between different models every single time — isn't that just annoying?

Then, just the other day, Hermes dropped MoA (Mixture of Agents), and it hit me: that ancient saying "three cobblers beat Zhuge Liang" had already spelled out where the AI world was heading this year.

MoA isn't some fancy new model — it's a brand-new collaboration paradigm

The core logic is simple: stop expecting one model to do everything, and instead, get multiple models working together.

You configure a "preset" that packs in a few reference models plus one aggregator model. When a tough problem comes up, the reference models run a round first, each offering their own analysis and suggestions, and then the aggregator model takes all that and produces the final answer.

For example: you're writing a tech analysis article. GPT-5.5 handles the smooth prose, DeepSeek V4 handles the logical reasoning, and Claude Opus does the final review and synthesis. Three cobblers each playing to their strengths — and the final output? Guess who ends up looking like Zhuge Liang?

Hermes also ran benchmarks, and the numbers speak for themselves

Hermes compared results on its own test set: Opus-4.8 going solo scored 0.7607, GPT-5.5 solo got 0.7412. But when they were grouped into a MoA preset (Opus as the aggregator, GPT-5.5 as a reference model), the score shot up to 0.8202 — a full 6 percentage points higher than the strongest single model, Opus.

This isn't just "averaging things out" — this is a real case of 1+1>2.

How to get started? It's a one-liner

The most straightforward way: just type this directly in the Hermes chat:

/moa 帮我把这篇产品需求文档拆成开发任务

That single command automatically switches to the default MoA preset, runs a full round, and then switches back to your original model. One sentence, zero configuration, zero learning curve.

For deeper usage, you can use /model default --provider moa to switch the entire current session into MoA mode, so every subsequent message is a multi-model collaborative output.

But where's the ceiling?

Honestly, MoA isn't magic.

Its biggest value isn't "making one model smarter" — it's letting the strengths of different models fully align within the same conversation round. When coding, use one reference model to gauge implementation complexity, another to review edge cases, and let the aggregator model make the final call — this kind of structured collaboration is far more reliable than "prompt engineering."

It does come at a cost: for each conversation turn, the reference models run a round, and then the aggregator runs another. Doubling your token consumption is par for the course. Officially, it supports up to 8 reference models in parallel, so theoretically you could have 9 models working for you at once — but whether your wallet can handle it is another story.

My take

The MoA direction matters more than most people realize.

Over the past year, the industry has been obsessed with "one model doing everything" — longer context windows, bigger parameters, higher benchmark scores. But MoA takes a completely opposite path: acknowledging that no model is all-powerful, yet using a combination of moves to compensate for individual models' weaknesses.

Isn't this the essence of an engineer's mindset? You don't wait for GPT-6 to descend from the heavens — just combine the best tools you already have, and the results are already impressive enough.

It's like cooking — you don't necessarily need one Michelin three-star chef to handle every dish. Let the Sichuan chef make the boiled fish, the Cantonese chef make the white-cut chicken, and the Shandong chef make the braised sea cucumber. Put it all on one table, and the flavors actually come out more authentic.

Go check which "cobblers" (model combos) you've got on hand, and give this "group chat" style collaboration a try — see if you can turn them into a "Zhuge Liang"? Drop a comment and let's talk.