Run Hermes free forever — a real AI employee on your own machine.

Here's the whole idea in one breath. Hermes is the agent — it plans, runs commands, writes files and finishes work on your Mac. Until now the brain behind it lived in someone else's cloud. LiquidAI just shipped LFM2.5 — a small, sharp model built for exactly one job: being an agent's brain on the device itself. Plug it into Hermes and the meter is gone. Forever.

A brain built to live in your pocket, not in a datacenter.

LiquidAI released LFM2.5-2.6B — two point six billion parameters, trained on around thirty-four trillion tokens, with a 128K context window and an open-weight license.

The pitch isn't "biggest model ever". It's the opposite: an agentic model small enough to run on phones, laptops and robots, tuned specifically to plan, call tools and work through multi-step tasks on the device itself.

Look at what their own benchmark table measures: ToolSandbox 77.83 (ahead of a model nearly four times its size), Multi-IF 80.07, IFStruct 85.49. Tool calling and instruction following. That's agent duty — and it predicts exactly what you'll see below.

"Best model for your local hardware 8GB" — trained with Hermes in the loop, able to navigate phones, computers and bots. When the Hermes people point at a local brain for Hermes, that's the one you wire in first.

Notice what those benchmarks measure: tool calling and instruction following. Not poetry, not giant codebases. Keep that in mind — it predicts exactly what my tests found.

The Zero-Meter Engine — three parts, one machine.

The routing rule that makes it work: the free brain is your default for everyday agent duty. Heavyweight builders stay one profile flag away — hermes -p <big-brain> — for the jobs that genuinely need one. You choose when a meter runs. It's never on by default.

And a meter changes how you use your agent more than you think.

When every message costs something, you ration. You batch your questions. You hesitate before letting an automation run every ten minutes. You think twice about cron jobs that fire all night.

The moment the brain is local, that voice in your head goes quiet.

Let it watch a folder all day. Let it summarise every session. Let the wake-word listen around the clock. Run it on a plane. Feed it documents you'd never send to a cloud. The habit changes before the tooling does — you stop asking "is this worth a call?" and start asking "why isn't this automated yet?"

Five commands. Zero accounts.

This assumes Hermes is installed. Everything else is below, and none of it asks for a card or a key.

# 1 · engine room (skip if you have Ollama)
brew install ollama
ollama serve

# 2 · pull the ACTUAL new model — the 2.6B from the announcement
ollama pull hf.co/LiquidAI/LFM2.5-2.6B-GGUF:Q4_K_M

# 3 · rebuild it with Ollama's native LFM renderer (the fix explained under this block)
cat > /tmp/Modelfile.lfm26 <<'EOF'
FROM hf.co/LiquidAI/LFM2.5-2.6B-GGUF:Q4_K_M
RENDERER lfm2
PARSER lfm2-thinking
PARAMETER repeat_penalty 1.05
PARAMETER temperature 0.2
PARAMETER top_k 80
EOF
ollama create lfm2.5-2.6b -f /tmp/Modelfile.lfm26

# 4 · give Hermes the profile
mkdir -p ~/.hermes/profiles/lfm
cat > ~/.hermes/profiles/lfm/config.yaml <<'EOF'
model:
  default: lfm2.5-2.6b
  provider: ollama
  base_url: http://127.0.0.1:11434/v1
  context_length: 65536
  ollama_num_ctx: 65536

toolsets:
  - hermes-cli
EOF

# 5 · verify before you trust it
hermes -p lfm status        # Model line must say lfm2.5-2.6b
hermes -p lfm -z "Reply with exactly: READY"

Why step 2 uses that long address: the short tag ollama pull lfm2.5 quietly gives you a different 8B variant — not the model from the announcement. And why step 3 exists: the direct download's chat template kills agent loops with a "2 or more assistant messages" error; the three-line rebuild swaps in Ollama's native LFM renderer and everything runs clean. One habit to keep: glance at its first file write — if you ever see a literal \n in the text, just tell it to rewrite the file.

From then on, every Hermes feature — chat, tools, skills, cron, the works — runs against a brain that costs nothing per message and never phones home.

Inside the Agent OS, free brain selected.

This is my actual dashboard. The Hermes tab's profile rail picked up the new lfm profile automatically — one click and the whole chat runs on the free local brain. I asked it what it is; the answer below came back in seconds, offline-capable, nothing metered.

The Agent OS · Hermes tab · lfm pill active — LFM introducing itself from inside the machine.

Now the part you actually came for — what this stack builds. The Zero-Meter Engine is one half of the routing rule; the other half is that every heavyweight engine in the Agent OS is one profile flag away. These are real builds from that routing — GoldieBench-class games, every one playable right now, every one playtested before it went on this page. I died in the shooter while testing it. That's how real they are.

Playable 3D game · built by the Fusion engine · GoldieBench 9.0

The Dragon Realm

A frozen open world with a wyrm circling the sky. WASD to roam, F to draw steel, click to swing.

Play it →

Playable 3D RPG · built by Fable 5 · GoldieBench 9.0

Frostvale

Sword, shield, moonlit peaks and a wolf-culling quest. A whole Skyrim-flavoured valley in one file.

Play it →

Playable FPS · built by Mixture of Agents · GoldieBench 8.6

DOOM, One File

A raycast shooter with enemies that shoot back — they took me down mid-playtest. Click to fight.

Play it →

Playable racer · built by GLM · GoldieBench 9.0

Neon Drive

An endless synthwave highway into the sun. CRT scanlines, wireframe hills, 83 km/h of 1984.

Play it →

And one more, built fresh while this page was being written — I asked the stack for a full SaaS landing page for the engine itself. One prompt to a builder profile, one factual fix round, playtested, done:

Full SaaS landing page · built via hermes -p qwen-3-8 · one prompt

The Zero-Meter Landing Page

Aurora hero, live typing terminal, glass feature cards, count-up stats, glowing pricing card, FAQ — a launch-ready page for the exact stack this guide wires up.

Open the page →

And the free brain's own contribution to this page: while those builds came from the heavyweight lane, the local swarm did the grunt work — four parallel LFM workers auditing this whole OS in 178 seconds, for nothing. Builders build, the free brain runs the errands. That's the routing rule working exactly as designed.

And the speed receipt, straight off the terminal:

I used to pay for every thought. Now the API is free.

Here's how this actually worked for me until now. Every single thing my agents did ran through a paid API. Every chat, every summary, every scheduled check — a metered call to someone else's cloud, on someone else's servers, with my documents riding along.

So I did what everyone does: I rationed. Automations I wanted got scaled back. Watchers ran hourly instead of constantly. Some ideas never shipped at all, because they'd have been "always on" — and always on meant always paying.

This model flips that. The same API my agents call is now a free one, running on my own Mac. Same Hermes, same tools, same skills — but the everyday layer costs nothing per message, works with the WiFi off, and nothing I feed it ever leaves the machine.

The ideas I used to turn down are just… running now. You can have the same thing set up tonight — it's the five commands above.

Real people. Real wins. Inside the Boardroom right now.

These are members' own posts as they happened — agency owners, ecom founders, course creators running this same stack.

Real member · built their first agent the same day

Real member · cut the monthly AI bill right down with free + local engines

Real member · whole workflow automated from one dashboard

What this gives you back.

What's actually stopping you isn't technical.

Do it properly — five steps.

The whole thing in 30 seconds.