Hermes Agent Best Auxiliary Models (2026)

Hermes agent best auxiliary models, answered properly: what works, what doesn’t, and what I’d do first.

Every /goal turn and /loop condition fires a judge, and every sub-agent burns tokens — none of it needs your frontier model. Here are the best auxiliary models for Hermes Agent, where to configure them, and the one place a cheap judge is a false economy.

What auxiliary models actually are

Your main model is the brain that does the work you asked for. Auxiliary models are the smaller brains Hermes uses for the jobs around that work — and configuring them is one of the quietest cost wins available.

The clearest documented example is the goal judge. When you run /goal, a judge decides after each turn whether the objective is met; when you run /loop --until, the same judge checks your condition after each wakeup. That judge is an auxiliary task with its own configurable provider and model — auxiliary.goal_judge.* in your config — separate from the brain doing the work.

The same logic extends to everything background-shaped: sub-agents doing grunt work, routine tool calls, memory lookups, summaries. None of it needs the model that plans your campaigns.

Why this matters: the judge runs constantly

A judge verdict fires after every turn of a goal and every tick of a conditional loop. Point that at a frontier model and you’re paying premium rates for a yes/no answer, dozens or hundreds of times per task.

Point it at a small fast model and the quality of the verdict barely changes while the cost of it nearly disappears. That’s the entire auxiliary game: frontier quality where it’s felt, small models where it isn’t.

The best auxiliary models, ranked by job

LFM2.5-2.6B — the default pick

2.6B parameters, 128k context, runs on ~8GB, and — the reason it tops this list — its final post-training stage was agentic RL run through the Hermes harness itself. In my testing it handled real tool calls fast without dragging the machine, and it’s exactly what I’d run as a sub-agent for grunt work.

Full breakdown in the LFM2.5-2.6B guide.

DeepSeek V4 Flash — for high-frequency steps

The lighter, faster V4 variant, priced for the hundreds of small steps an agent loop generates — reading files, choosing tools, formatting, follow-ups. The routing pattern is in the Flash guide.

Ling 3.0 Flash — the free tier

124B MoE with ~5B active, free three separate ways, and every functional build I tested actually worked. As a free auxiliary brain for judging and routine calls it’s hard to argue with. Detail in the Ling guide.

A small local Qwen or Gemma — for private pipelines

When the auxiliary work touches client data — summarising documents, memory operations — a small local model keeps every one of those background calls on your machine.

What not to do

• Don’t use Maple Preview or LFM for the main coding brain — both are honest misfits there. Auxiliary is where they shine.
• Don’t let an auxiliary model judge high-stakes output. A cheap judge on “is the queue empty” is free money; a cheap judge on “is this client deliverable good” is a false economy.
• Don’t skip the tool-call test. Small models fail by fumbling a tool call, not by sounding wrong — test the thing that actually breaks.
• Don’t forget the judge fails open. A broken judge never wedges your loop — but it also means a misconfigured auxiliary can silently stop enforcing your condition. Check it fired.

Setting it up

1. Override the judge in ~/.hermes/config.yaml — auxiliary.goal_judge.provider, .model and .max_tokens. The same overrides apply to /loop --until conditions.
2. Run sub-agents on a small model profile, and keep the frontier brain for the parent.
3. Route the choice by data sensitivity: free tiers for public work, local for private.
4. Watch one day’s usage after the change — the drop in background spend is the confirmation it’s working.

FAQ

What are the best auxiliary models for Hermes Agent?

LFM2.5-2.6B as the default — it was trained through the Hermes harness — with DeepSeek V4 Flash for high-frequency steps, Ling 3.0 Flash as the free option, and a small local Qwen or Gemma for private work.

What is an auxiliary model in Hermes?

A smaller model handling background tasks around your main work — most visibly the goal judge that checks /goal progress and /loop –until conditions, configurable via auxiliary.goal_judge.* in your config.

Why not use one model for everything?

Because the judge and sub-agent calls fire constantly. Paying frontier rates for yes/no verdicts is the quietest leak in an agent budget.

Does a cheap judge give worse verdicts?

On simple conditions — is the build green, is the queue empty — barely. Keep expensive judgement for genuinely high-stakes evaluation.

Where do I configure it?

In ~/.hermes/config.yaml under auxiliary.goal_judge — provider, model and max_tokens — and the same settings govern loop conditions.

Can auxiliary models run locally?

Yes, and for anything touching client data they should. Small local models keep every background call on your machine.

How do I know a small model is up to it?

Test tool calls, not conversation. Small models fail at calling tools correctly, not at chatting.

What happens if the judge breaks?

It fails open — the loop never wedges. Which also means you should confirm it’s actually firing, or your condition isn’t being enforced.

The bottom line

The best auxiliary models for Hermes Agent are small, fast and cheap — LFM2.5-2.6B first, V4 Flash for volume, Ling free, local for private — because the judge and the grunt work fire constantly and never needed a frontier brain. Route them once in config, test a tool call, and keep the expensive model for the decisions that are actually felt.

About Julian Goldie

I run Goldie Agency, a 7-figure SEO agency, and teach this daily on a 400K+ subscriber YouTube channel. 240+ client projects on Upwork at a 100% job-success score, 10+ years through every major Google update. My systems are in the AI Profit Boardroom; my link building book is free here.

Related reading

• LFM2.5-2.6B
• DeepSeek V4 flash harness
• Ling 3.0 Flash

Last updated August 2026. This is the living guide to Hermes agent best auxiliary models — it gets updated as the tools change.