Nous Research has released Hermes Mixture of Agents 2.0 (MoA 2.0), an update to its open-source Hermes Agent framework that lets users combine several AI models into a single system which, the company says, outperforms today's strongest publicly available models, including Claude Opus 4.8 and GPT-5.5. The pitch is pointed: you no longer need access to one restricted frontier model if a mix of accessible ones can beat it.

What is Mixture of Agents 2.0?

MoA is not a new model but a multi-model architecture. A user configures a preset made of one or more "reference models" plus a single "aggregator." The reference models each analyze the request independently, then the aggregator reads all of their outputs, synthesizes a final answer, and handles any tool calls.

MoA already existed in Hermes as a mode you toggled. The main change in 2.0 is that each named preset now appears as a selectable "virtual model" in the model picker, listed right alongside Claude, GPT and Grok, across the CLI, the desktop client, and gateways like Telegram and Discord. You pick a preset like any other model and Hermes routes the request through the ensemble. A /moa [prompt] command also allows a single one-shot call that reverts to your normal model afterward, for use only on demanding tasks.

How can a mix of models beat the best single one?

The idea is an expert panel. Each model brings different strengths and blind spots, so when several analyze the same problem independently, their mistakes tend not to overlap. The aggregator then acts as the chair, weighing the perspectives and composing an answer stronger than any single member would produce alone. It is the same principle behind ensemble methods in machine learning, applied at the level of entire models rather than smaller components, which is why a good panel can beat its most capable individual member.

What do the benchmarks show?

Nous points to its upcoming HermesBench. A default preset using GPT-5.5 and DeepSeek as reference models and Claude Opus 4.8 as the aggregator scored 0.8202, versus 0.7607 for Opus 4.8 alone and 0.7412 for GPT-5.5 alone, roughly 8% above Opus and 11% above GPT-5.5. Nous chief engineer Teknium said on X that the team is still testing open-source combinations aiming for "Opus-level output at much lower cost."

The important caveat is transparency. HermesBench is not yet fully public, so these are Nous Research's internal results, and a complete public leaderboard is still in preparation. Readers should treat the numbers as the company's own claims until the benchmark and its methodology are released.

What are the engineering details?

Several design choices are worth noting for anyone deploying it. Prompt caching is preserved by appending reference outputs to the end of the latest user turn rather than inserting them mid-history, which keeps a stable context prefix hitting the cache and holds down cost. Nested MoA is banned: an aggregator cannot itself be another preset, which blocks runaway recursive mixing and the cost and debugging headaches that come with it. Reasoning is transparent, with each reference model's full output shown as its own labeled block so you can read GPT, Claude and Grok separately before the aggregator streams the final answer. And full tool access is reserved for the aggregator, while reference models receive a simplified conversation stripped of system prompts and tool history, to cut cost and avoid provider-level refusals from stricter services.

Why release this now?

The timing is not accidental. On June 12, a US export-control directive forced Anthropic to suspend Fable 5 and Mythos 5 for all users worldwide, a ban only lifted on July 1 after a 19-day shutdown, while access to the leading models has generally grown more expensive and rate-limited. Nous built its messaging around that reality. In its announcement it wrote that "the strongest models are access-restricted, available to only a few," and positioned MoA 2.0 as the alternative: rather than depend on a single restricted super-model, assemble the accessible ones into a system that outperforms them.

What is the catch?

MoA is expensive. Each call multiplies token usage by roughly the number of reference models, so Nous recommends reserving it for "the 10% of tasks that most need quality" rather than every interaction. Teknium says the cost problem should ease sharply once open-source model combinations can match the quality of the proprietary ones, removing the need to route through pricey frontier APIs at all. MoA 2.0 shipped as a core feature of Hermes Agent v0.17.0 on June 19 and was refined in the July 1 "Judgment Release," v0.18.0, with more detailed trace persistence and security hardening.

Frequently Asked Questions

What is Hermes MoA 2.0?

Hermes Mixture of Agents 2.0 is an update to Nous Research's open-source Hermes Agent framework that lets users combine multiple AI models into a single system. Several "reference models" analyze a request independently, then an "aggregator" model synthesizes a final answer, and the preset appears as a selectable virtual model.

Does combining models really beat a single frontier model?

Nous says yes, citing its HermesBench where a preset scored about 8% above Claude Opus 4.8 and 11% above GPT-5.5. However, HermesBench is not yet public, so these are the company's internal results, and independent verification is not yet available.

Why would I use MoA instead of one model?

It can produce higher-quality answers by pooling several models' strengths, and it reduces dependence on any single provider, which matters as top models face access restrictions and rising costs. The tradeoff is much higher token usage per call.

How much does MoA cost to run?

Each call uses roughly as many times more tokens as there are reference models, so it is significantly more expensive than a single model. Nous recommends using it only for the most quality-sensitive tasks, and expects costs to fall as open-source combinations improve.