"The strongest models are gated, and access is granted only to a select few. Hermes Agent now exposes MoA presets as virtual models, giving you capabilities beyond the publicly available frontier."
— Nous Research (@NousResearch)
✦
I · catch-up in 30 seconds
What Nous just shipped.
Nous Research dropped Mixture of Agents 2.0 inside Hermes Agent.
Here's the plain-English version.
You pick a few models — any providers, mixed together. Opus, GPT-5.5, an open model, whatever.
You save that mix as a preset. And Hermes treats the whole mix as one model.
Under the hood, the other models do private analysis first. Then one "aggregator" model reads all of it and writes the final answer — and runs your tools.
You flip between mixes mid-chat with /moa. And it keeps prompt caching, so it stays fast.
The result Nous is claiming: 8% higher than Opus 4.8, and 11% higher than GPT-5.5.
Not a new model. A new system built on the models we already have.
Thinking it? "8% and 11% — says who?"
Those are Nous's numbers on HermesBench, their own benchmark, which isn't public yet.
So treat it as their claim for now. But the idea behind it is already proven — I'll show you my own results below.
II · from the team
The post that kicked it off.
Here's the announcement straight from Nous. Read their framing closely — it's the whole point.
The strongest models are gated. Only a select few get them.
MoA flips that. You assemble your own mix that sits above the public frontier — without waiting for anyone to grant you access.
the announcement ↓
Heads up: live X embeds need a connection to load. Offline they show as a plain link — that's normal.
III · how it works
A panel of experts beats one genius.
Picture one brilliant person answering a hard question alone.
Now picture a panel of brilliant people — each writes their own take, privately — and a sharp chair reads all of them and gives you the best combined answer.
The panel wins. Every time.
That's Mixture of Agents.
The reference models are the panel. The aggregator is the chair.
One question goes in. Several models think. One clean answer comes out — better than any single one of them could give.
1 prompt → many models think → 1 answer, better than any of them alone. That's the whole trick.
✦
IV · the big idea
Stop chasing the model. Build the system.
Everyone is waiting on the next model.
The next Opus. The next GPT. The thing that'll finally change everything.
But look at what just happened: a mix of today's models beat the best single model on the market.
No new release. No gated access. Just a smarter way to use what's already here.
That's the lesson MoA hands you for free: the model is not the moat. The system around it is.
The model is a part you can swap. The system is the thing you own.
And the timing makes this huge right now.
Fable 5 is in preview. GPT-5.6 is delayed. The next frontier models aren't really in your hands yet.
So the winning move isn't waiting. It's squeezing more out of the models you already have — by combining them.
Thinking it? "I'll just wait for the next big model."
You'll wait months — and then you'll still be using one model at a time.
A system that mixes models beats a single model today, and it'll beat the next one too. The system compounds. The model expires.
Models rotate through. The system stays — and keeps you ahead of whatever ships next.
✦
V · proof it works
I've been running this for weeks.
MoA isn't a one-off trick. It's a pattern. And I've already built two systems on the exact same idea — a panel of models, fused into one answer.
my system · panel engine
Fusion
A boardroom of models. One prompt goes to a panel — Fable 5, GPT-5.5, GLM, Kimi — and a judge fuses the best of all of them into one reply.
Beats Fable 5 — at half the price.
my system · panel engine
Sakana (Fugu)
The same panel idea, a different recipe. Several models answer, the system picks and merges the strongest — one answer out, for a few dollars a run.
Frontier quality, tiny spend.
now · from Nous
Mixture of Agents 2.0
The same idea, now baked into Hermes as a virtual model. Reference models analyse, an aggregator answers.
Beats Opus + GPT (per Nous).
Three names. One pattern. Mix models, fuse the best, beat any single one.
When my Fusion stack beat Fable 5 at half the cost, that wasn't luck. It's the same physics behind MoA. The system wins.
Thinking it? "This only works for big AI labs."
I'm one person and I run all three. Fusion and Sakana are wired into my Agent OS, and MoA is one command in Hermes.
If you've got the system, you've got the edge — lab or not.
Three names. One pattern. Whatever the label, it's a panel of models fused into one better answer.
✦
my own result · Fusion vs a single frontier model
Fusion (a panel of models)wins · half the price
Claude Fable 5 (one model)2× the cost
My Fusion stack out-scored Fable 5 — at roughly half the spend. Same lesson as MoA: the system beats the model. (My own test, in the Fusion guide.)
✦
VI · old way vs new way
Why this changes your move.
Chasing the model
always behind
Pick one model and hope it's the best
Hit that model's ceiling — and stop there
Wait months for the next release to save you
Beg for access to the gated frontier models
Pay top dollar for one premium model
Building the system
beyond the ceiling
Mix several models into one virtual model
Beat the best single model — today, with MoA
No waiting — squeeze more from what's already here
Go beyond the public frontier, no access needed
Hit frontier quality for a fraction of the cost
Thinking it? "Sounds complicated to run."
In Hermes it's one command — /moa — to switch your mix mid-chat. Presets do the rest.
And in the Agent OS, MoA, Fusion and Sakana are already wired in. You just pick your panel.
Get the system
Hermes Agent — with the panel built in.
MoA, Fusion and Sakana all run on the same idea, and they all live in Hermes Agent inside the Agent Operating System — the dashboard I run my whole business on.
The full Agent OS zip — Hermes, MoA, Fusion + Sakana wired in
Coaching calls where I build a model panel with you, live
Daily tutorials so you're stacking models, not just reading about them
A room of 3,600+ operators running this exact stack
The prompts, the presets, and a member map for your city
I'm not reporting on MoA from the sidelines. I run Fusion and Sakana every day, I tested them against the frontier myself, and MoA slots straight into the same setup.
0K YouTube subscribers
0+ members in AIPB
0countries · live members
0K followers on X
I'm not going to paste invented quotes here. The wins are real and written by the members themselves — read them in their own words.
HermesBench isn't public yet, so the exact 8% and 11% are Nous's claim for now. Fair to stay a little careful on the numbers.
But the pattern is not in doubt. I've run it myself — Fusion beat Fable 5 at half the cost.
The people who win the next year won't be the ones with the best model. Everyone gets the same models eventually.
They'll be the ones with the best system — the one that mixes, fuses, and gets more out of every model that ships.
So stop refreshing the launch pages. Start building the panel. The model expires. The system compounds.
the recap
What you now know.
i.
MoA 2.0 shipped in Hermes. Stack any models into one virtual model.
ii.
It beats the frontier. +8% on Opus 4.8, +11% on GPT-5.5 (per Nous).
iii.
Panel beats genius. Reference models think, an aggregator fuses the best.
iv.
Same idea as Fusion + Sakana. My Fusion stack beat Fable 5 at half the cost.
v.
Perfect timing. Fable 5 + GPT-5.6 are in preview — combine today's models instead.
vi.
The model doesn't matter. The system does. Build the system.
The model expires. The system compounds.
Build the panel
Stack models. Beat the frontier. Own the system.
Grab Hermes Agent inside the Agent Operating System in the AI Profit Boardroom. MoA, Fusion and Sakana — the whole panel — wired and ready. You pick your models, it does the rest.
The full zip. Hermes + MoA + Fusion + Sakana presets.
Coaching calls where we build your model panel together
