Nous Research released Hermes Mixture of Agents 2.0 on Sunday, which combines outputs from multiple large language models, including GPT, Claude, and DeepSeek, to produce responses that outscore any individual model on standard benchmarks.
According to a report, the MoA 2.0 is an update to Nous Research's existing Hermes Agent framework and preserves its open-source structure.
How The System Works
Hermes MoA 2.0 operates as an ensemble layer. It queries several base models in parallel, collects their outputs, and synthesizes a final response. The approach, known as Mixture of Agents, treats distinct AI models as specialist contributors rather than requiring a single model to handle every task alone.
Users can configure which models participate in a given ensemble. The default configuration draws on GPT, Claude, and DeepSeek, three models that represent different training philosophies and data compositions. By pooling their outputs, MoA 2.0 captures complementary strengths.
Benchmark results shared with the release show MoA 2.0 outperforming each component model individually across reasoning, coding, and instruction-following tasks. The margin is meaningful on long-horizon reasoning tests, where single models often lose coherence.
The framework remains open-source, meaning researchers and developers can inspect the architecture, swap out base models, and adapt the ensemble for specific use cases.
Also Read: Anthropic Tops OpenAI With $965B Valuation As AI Funding Race Resets
Open-Weight Labs Push Into Agent Orchestration
Nous Research has built a reputation for open-weight model releases aimed at the research community. The original Hermes Agent framework established a baseline for multi-model orchestration earlier in 2026.
The broader context is an accelerating open-weight AI development cycle. Z.ai published GLM-5.2 in early July 2026, positioning it as an open-weight coding model for long-horizon engineering tasks. The release follows a pattern of open-weight labs targeting specific capability domains where closed models hold reputational advantages.
Also Read: Experts Say OpenAI's 5% Stake Would Push AI Deeper Into State Control
Qwen's former technical lead Junyang Lin publicly argued in late June 2026 that agentic systems represent the correct next step for AI development. That argument aligns with the design philosophy behind MoA 2.0, which treats agents and model combinations as a path toward capability gains that individual training runs cannot easily replicate.
The Hermes MoA release also arrives amid active debate in the AI research community about the correct role of foundation models versus agent layers.
Andrej Karpathy cautioned earlier this week that agent-first development risks repeating mistakes from OpenAI's earlier research cycles. Nous Research's approach attempts a middle path, using strong foundation models as inputs while adding an orchestration layer on top.
Also Read: OpenAI Eyes $1 Trillion IPO While Microsoft Holds The Biggest Prize
What To Watch
Hermes MoA 2.0 has not yet been tested against the most recently released frontier models. Claude Sonnet 5 and updated GPT variants released in mid-2026 may alter the benchmark picture. Nous Research has not published a formal academic paper alongside the release.
The practical significance for developers is clear. An open-source tool that demonstrably improves on closed model benchmarks by combining them lowers the barrier for research teams to access top-tier reasoning capability without paying frontier model API costs for every inference call.
For the AI industry, MoA 2.0 adds weight to the argument that model diversity, rather than a single dominant model, may define the next phase of AI deployment. Watch for responses from OpenAI and Anthropic on ensemble-based approaches in the months ahead.
Read Next: The Two-Faced Trump Anthropic Policy The White House Won't Explain