/Tech7h ago

Nous Research releases Mixture-of-Agents presets for Hermes Agent, beating GPT 5.5 by 11% on HermesBench

Story Overview

Nous Research has added Mixture-of-Agents presets to Hermes Agent, letting users run several frontier models in parallel with an aggregator that handles the final output and tool calls. The setup delivers the reported gains on the company's internal HermesBench, where one preset reached 0.8202 versus 0.7607 for standalone Opus 4.8 and 0.7412 for GPT-5.5.

671852513929.9K

#331

Original post

DeFi狙击手 | Ai🕊️@bi_9527zx

最强模型都被大公司锁着，只有少数人能用。

Hermes Agent 现在把 MoA 做成了虚拟模型，在他们 HermesBench 上比 Opus 4.8 强 8%、比 GPT 5.5 强 11%。

原理很简单：先让多个顶级参考模型并行思考（只给对话内容，不给工具），再由一个 aggregator（汇总模型）综合它们的结果去输出答案并调用工具。

不是自己从头训新大模型，而是把现有强模型聪明地组合起来用。

想自己体验？装个 Hermes Agent 后直接输入 /moa 就能切换（推荐先跑 hermes setup --portal 一键搞定）。

普通人也能用到这种 frontier 级能力，开源社区这波操作真的香！

感谢@NousResearch团队持续推进 🚀

#HermesAgent #MoA #开源AI

Nous Research@NousResearch

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: 8% higher than Opus 4.8 and 11% higher than GPT 5.5 on our upcoming benchmark.

9:30 PM · Jun 26, 2026 · 29.2K Views

Developer Impact

How the presets fit into daily agent use

Users pick the new moa virtual provider the same way they select any other model, via CLI, desktop, or dashboard, and the feature keeps full tool loops and prompt caching intact while billing flows through the underlying providers.

Open Question

Benchmarks that still need external confirmation

All numbers come from Nous's own HermesBench, described as coming soon to a public leaderboard, so the exact lift on other widely used evaluations remains an open variable for now.

Sentiment

Many users praised the open-source Hermes Agent MoA for outperforming Opus and GPT models on benchmarks because it delivers strong performance and makes advanced AI practical and accessible.

Pos

97.4%

Neg

2.6%

38 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

VIEWS751BOOKMARKS3LIKES10RETWEETS5REPLIES1

Connor Shorten@CShorten30

Mixture-of-Agents! 🧑‍🍳

From @NousResearch: ===== "Mixture of Agents combines the models you can access into one. [...] The reference models answer in parallel. Your aggregator reads their responses and runs the task, calling tools and repeating the loop as needed." =====

I think this looks really awesome! I am super excited about the multi model systems we are building at Weaviate. 🧬

This emerging sample-and-judge pattern is incredibly powerful. I also like Perplexity's Model Council as an example of this. 🏟️

OpenRouter is an interesting counterpoint, instead of sampling every model and judging, you route to the right one upfront and only pay for one inference. Predict the winner vs. run them all. 🎱

And the classic inference ensemble is still undefeated. We're huge fans of Hybrid Search at Weaviate, fusing scorers with complementary failure modes is a similar idea at the retrieval layer. 🛠️

Congrats to @NousResearch on MoA! 🚀

Nous Research@NousResearch

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: 8% higher than Opus 4.8 and 11% higher than GPT 5.5 on our upcoming benchmark.

3h751103