GoldieBench Blog · 2 min read
The Best Local Model For Hermes Agent — Decided By 45 Real Builds
We ran every runnable local model through 45 one-shot builds and playtests. The best local model for Hermes Agent in 2026, with real 0–10 scores, tok/s numbers, and live demos you can click.
Hermes Agent is the body: always on, on your machine, running tools.
The local model is the brain — and picking the wrong one wastes your whole setup.
This isn't a vibes list: every model below went through the same GoldieBench gauntlet (45 one-shot builds, real rendered screenshots, an Opus 4.8 vision judge scoring 0–10) on a 36GB Mac.
The short answer
Best overall quality: Qwable 5 27B Coder — 7.14 average across the board, the only local model that one-shots rich 3D scenes.
Trade-off: .20 tok/s, so agent loops feel slow.
Best speed-to-quality (our pick for Hermes): Agents-A1 — a 35B MoE with .3B active, agent-tuned by InternScience. 4.83 average, 3 local golds, and it runs at 95 tok/s — nearly 5× Qwable's speed.
For long agentic loops (tool calls, file edits, multi-turn plans) that speed is the difference between a snappy agent and a coffee break.
Best lightweight daily driver: Gemma 4 12B (MLX) — 3.98 average, .62 tok/s with Ollama's MLX+MTP path, tiny footprint.
This is the model from Julian's "Run Hermes Free Forever" video: pair it as a sub-agent under a stronger main model and your token bill for auxiliary tasks drops to zero.
The local leaderboard, verified
Full table with clickable demos lives at the local models board.
Highlights: Qwable 7.14 · Agents-A1 4.83 · Gemma 4 MLX 3.98 · Laguna XS 2.1 3.93 · Qwythos 2.98.
Every score links to the actual build the judge saw — click in and play them.
How to wire it into Hermes
Two commands with Ollama, then point Hermes at it: ollama pull the model, set it as the provider in your Hermes profile (provider: ollama, base_url: http://localhost:11434/v1).
The full walkthrough with profiles and agentic workflow examples is in the Hermes + Gemma 4 guide on agentos.guide.
Main model vs sub-agent: the setup that saves tokens
The winning pattern from our testing: a frontier model (Claude, GPT-5.6) as the main planner, and the local model as the sub-agent for grunt work — summaries, file reads, boilerplate.
Gemma 4 or Agents-A1 handle those at $0, and your paid tokens only touch the hard reasoning.
FAQ
What is the best local model for Hermes Agent in 2026?
For agent loops we pick Agents-A1 (35B MoE, 95 tok/s, agent-tuned, 4.83 GoldieBench average). For maximum one-shot build quality pick Qwable 5 27B (7.14). For a lightweight free sub-agent pick Gemma 4 12B MLX.
How much RAM do these local models need?
Gemma 4 12B runs comfortably on 16GB. Agents-A1's Q4 build is 21GB and wants a 32GB+ Mac. Qwable 5 27B similar. All scores here came from a 36GB MacBook.
Are these scores real or vendor benchmarks?
Real. Every model ran the same 45 one-shot builds; each build was rendered, screenshotted and scored 0–10 by an Opus 4.8 vision judge. The demos are public — click any score on the local board and play the build.
Can I run Hermes Agent fully offline with these?
Yes — that's the point. Ollama serves the model locally, Hermes talks to localhost, and the whole agent loop works with no internet and no API bill.
Keep reading
Hermes Herald AI Agent: Which Brain To Run It On, With Real GoldieBench Scores
GEO Examples: Four Real AI Overview Wins, The Three Moves Behind Them, And The Models Writing The Answers
Best Generative Engine Optimization Agency: Seven Picks By Region, And How The Models Behind AI Answers Score🌐 Sister-site take: read this on agentos.guide