c35d8e0736
- raw: raw/xpost/2026-06-28-hermes-moa-vaibhavsisinty.md (NEW) - wiki: tools/hermes-desktop.md (updated — new MoA section: architecture, performance, composability, criticism, OpenRouter Fusion comparison) - wiki: index.md (updated — 38th update, Hermes Desktop row updated, new raw source) - wiki: log.md (new entry) Source: @VaibhavSisinty X post, shared by Pit Weber in OME Topic 3770 (Hermès Agents), 2026-06-28
49 lines
No EOL
2.3 KiB
Markdown
49 lines
No EOL
2.3 KiB
Markdown
---
|
|
type: xpost
|
|
source_url: https://x.com/vaibhavsisinty/status/2070741416649850898
|
|
retrieved: 2026-06-28
|
|
author: "@VaibhavSisinty"
|
|
is_thread: false
|
|
tags: [hermes, mixture-of-agents, moa, model-fusion, ensemble, nous-research, teknium, lambdua]
|
|
---
|
|
|
|
# Hermes "Mixture of Agents" (MoA) — Merge Any Models Into One Virtual Model
|
|
|
|
**Geteilt von:** Pit Weber in OME-Gruppe Topic "Hermès" Agents (Topic 3770), 2026-06-28
|
|
|
|
## Source: X Post
|
|
|
|
**Author:** Vaibhav Sisinty (@VaibhavSisinty)
|
|
**Posted:** 2026-06-28
|
|
**URL:** https://x.com/vaibhavsisinty/status/2070741416649850898
|
|
|
|
## Content Summary
|
|
|
|
Hermes "Mixture of Agents" (MoA) feature:
|
|
|
|
- **Merge any two (or more) AI models into one virtual model**
|
|
- One model runs as **reference**, one as **aggregator**
|
|
- Both run **in parallel** per task; aggregator synthesizes final output and handles tool calls
|
|
- Virtual model appears as a **single selectable model** in the model picker
|
|
- **Results:** 8% above Opus 4.8 solo, 11% above GPT-5.5 solo on hard agentic tasks
|
|
- Full Hermes features work **untouched**: Memory, tool use, skills, long sessions, cross-channel messaging
|
|
- **Any provider mix works:** OpenAI, Anthropic, OpenRouter, local models
|
|
|
|
### Community Clarifications
|
|
|
|
- **@Teknium** clarified: any number of models, not just two — even multiples of the same model
|
|
- **Criticism:** increased token costs; @lambdua called it "toy stage"
|
|
|
|
## Key Points
|
|
|
|
1. **Architectural approach:** Reference model generates draft responses; aggregator model synthesizes final output from reference + its own reasoning. Tool calls handled by aggregator.
|
|
2. **Performance gains:** Double-digit improvements on hard agentic tasks over solo frontier models (Opus 4.8, GPT-5.5)
|
|
3. **Transparency:** Virtual model is selectable in picker like any single model — no workflow change needed
|
|
4. **Composability:** Any provider combination (OpenAI + Anthropic, OpenRouter + local, etc.)
|
|
5. **Scalability:** Not limited to 2 models — N models supported, including same-model multiples (self-fusion pattern)
|
|
|
|
## Cross-References
|
|
|
|
- [[../../wiki/tools/hermes-desktop.md]] — Hermes Desktop main page
|
|
- [[../../wiki/concepts/llm/llm-model-fusion-ensembles.md]] — OpenRouter Fusion / DRACO-Benchmark (related ensemble approach)
|
|
- [[../../wiki/concepts/llm/ai-intelligence-commoditization-thesis.md]] — Intelligence commoditization context |