LFM2.5-2.6B might be the best free local model you can put behind Hermes right now — because its final training stage ran through the Hermes agent harness itself. Here’s how it performed in my own testing.
Short answer
- 2.6B parameters, 34 trillion tokens, 128k context, runs on around 8GB.
- Final post-training stage was agentic RL through the Hermes harness.
- Fast, private, offline — and it didn’t slow my machine down like other local models.
- Use it for agent work and tool calls. Don’t use it for coding.
Why this model is different
Most local models are trained on generic chat and then people try to bolt an agent onto them. LFM2.5-2.6B was post-trained the other way round.
Its final post-training stage was multi-turn agentic reinforcement learning run through the Hermes agent harness. In plain terms: they trained it inside the actual agent people use, not just on conversation data.
I’ll be transparent — the four post-training stages are listed as supervised fine-tuning, expert specialisation, multi-domain on-policy distillation, and agentic RL. I’m not going to pretend I can explain each one properly. What I can tell you is the practical effect, which is that it behaves like something built for agent work.
The specs
| Spec | Detail |
|---|---|
| Parameters | 2.6 billion |
| Training data | 34 trillion tokens |
| Context window | 128k tokens |
| Hardware | Runs on around 8GB |
| Built for | On-device and ultra-fast server-side deployment |
| Runs via | Hugging Face or LM Studio |
The headline is the memory footprint. This runs on ordinary hardware, and it can navigate phones and computers, not just answer questions.
How it actually performed for me
From what I’m seeing, it’s faster than Gemma and better than Gemma. That’s my own testing, not a benchmark chart.
The bigger deal is what it didn’t do. When I run a local model with Hermes, it usually slows my entire setup to a ridiculous crawl. This one didn’t. It was quick and it left the rest of my machine alone.
I tested it on real tool use rather than vibes. I gave it a skill command and a guide to learn from. It searched the web using the built-in Firecrawl API, read through the guide, created a skill locally from it, and reported the skill successfully learned. That’s a genuine agentic loop running on a free local model.
Then I pointed it at my memory: check my Obsidian vault and give me SEO keyword ideas for future content. It came back quickly with properly personalised keywords pulled from my own notes — things like agent OS setup guides and integration guides based on what’s actually in my vault. Not generic suggestions.
What I would and wouldn’t use it for
I wouldn’t use this for coding tasks. I want to be straight about that. It’s not what it’s designed for.
What it’s for is agent work — tool calls, memory retrieval, routing, and the high-volume jobs that eat tokens without needing a frontier brain.
| Use it for | Use something else for |
|---|---|
| Tool calling and skill execution | Serious coding work |
| Memory lookups against your vault | Long creative writing |
| Running as a sub-agent doing grunt work | Complex multi-step reasoning |
| High-volume, token-heavy routine tasks | Anything you’d bet the client on |
The pattern I like most is delegation. Use a frontier model as the brain of the engine, then hand the time-consuming non-frontier subtasks down to LFM. You get the quality where it counts and stop burning tokens where it doesn’t.
Old way versus new way
| Before | With a local model like this |
|---|---|
| Every message needed a paid API key | It runs free on your machine |
| All your data went to the cloud | Nothing leaves your machine |
| Local models slowed everything down | Fast enough to actually use |
| No Wi-Fi meant no agent | Works on a plane, offline |
| You rationed every token | You stop counting |
People sometimes tell me this isn’t really free. It is. Hermes is open source and free, and LFM is a free local model. The setup is a few commands.
Getting it running
- Pull the model through Hugging Face or LM Studio — either works.
- Point Hermes at it as your local model. See the local model setup guide for the wiring.
- Wire in your memory vault so it can answer from your own context. See best memory for Hermes Agent.
- Set it up as a sub-agent for the heavy routine work, and keep a frontier model for the hard calls.
- Test it on your own tool-use tasks before trusting it with anything that matters.
Worth pairing with Maple Preview if you want a fast local builder alongside it — two local models can cover most of an agent OS between them.
Want the full local setup? The Agent OS with free local models, multiple agent profiles and Hermes training is inside the AI Profit Boardroom, including a full course on running an agentic OS entirely free.
FAQ
What is LFM2.5-2.6B?
A 2.6 billion parameter local model from Liquid AI, trained on 34 trillion tokens with a 128k context window, designed for on-device and fast server-side agent work.
Why does it work well with Hermes?
Because its final post-training stage was multi-turn agentic reinforcement learning run through the Hermes agent harness. It was trained inside the agent, not just on chat data.
What hardware do I need?
It runs on around 8GB, so ordinary consumer hardware is enough. That’s a big part of the appeal.
Is it good for coding?
No, and I wouldn’t use it for that. It’s built for agent tasks — tool calls, memory retrieval and routine work.
Can it use tools and skills?
Yes. In my testing it searched the web through the built-in Firecrawl API, learned from a guide and created a skill locally, all on a free local model.
Does it work with an Obsidian memory vault?
Yes. I asked it to check my vault and suggest keywords, and it came back fast with genuinely personalised suggestions from my own notes.
Is it actually free?
Yes. Hermes is open source and free, and the model runs locally at no cost. Setup is a few commands via Hugging Face or LM Studio.
Can I use it alongside a frontier model?
That’s the best way to run it. Keep a frontier model as the brain and delegate the token-heavy routine subtasks to LFM.
The bottom line
Hermes supplies the hands and LFM2.5-2.6B supplies a brain that was actually trained for the job — 2.6 billion parameters, 128k context, running free and private on your own machine. Don’t use it for coding. Do use it for agent work, and it’s one of the best free local models you can put behind Hermes right now.
About Julian Goldie
I run Goldie Agency, a 7-figure SEO agency, and teach this stuff daily on a 394K+ subscriber YouTube channel. I’ve delivered 240+ client projects on Upwork at a 100% job-success score over 10+ years of ranking sites through every major Google update. The systems I actually run are inside the AI Profit Boardroom, and my link building book is free here.
