The best open source models for Hermes Agent aren’t the biggest ones — they’re the ones matched to the job. Here are the four I’d actually run, what each is genuinely good at, and the difference between open source, free and local that trips most people up.
Short answer
- LFM2.5-2.6B — best for agent work; trained through the Hermes harness itself.
- Ling 3.0 Flash — best all-rounder; 124B MoE with ~5B active.
- Maple Preview — fastest; MIT licensed, but weak at coding.
- Open source, free and local are three different things — don’t confuse them.
Open source, free and local are three different things
People use these words interchangeably and then pick the wrong model. They’re not the same thing.
| Term | What it actually means |
|---|---|
| Open source | The weights are published. You can self-host, inspect and fine-tune it. |
| Free | It costs nothing to use right now — which includes free tiers of closed models. |
| Local | It runs on your own hardware, with nothing leaving your machine. |
An open-source model can be run locally or through somebody’s paid API. A free model isn’t necessarily open source. And a local model has to be open source, but not every open-source model is small enough to run locally.
This page is about the first column: open weights you can actually own. If you specifically want no-cost, see the best free models; if you want it on your own machine, see the best local models.
The best open source models for Hermes Agent
LFM2.5-2.6B — the one built for the job
2.6 billion parameters, 34 trillion training tokens, a 128k context window, and it runs on around 8GB. What makes it the top pick for Hermes specifically is that its final post-training stage was agentic reinforcement learning run through the Hermes harness itself.
In my testing it was faster and better than Gemma, and crucially it didn’t drag my whole machine down the way local models usually do. It handled real tool calls — searched the web, learned a skill, read my memory vault and came back with personalised keywords.
Use it for: tool calls, memory lookups, routing, high-volume routine agent work. Don’t use it for: coding. Full detail in the LFM2.5-2.6B guide.
Ling 3.0 Flash — the best all-rounder
A 124-billion-parameter mixture-of-experts with roughly 5 billion active per token, and a 250K context window. That gap is why it’s quick despite the size.
Every one-shot build I tested actually worked — an agency site, a habit tracker, a calorie tracker. As a free brain for Hermes it’s excellent: fast replies, quick skill learning, and it connects to MCPs.
Weakness: front-end design is plain. Fix it with a design skill or hand the polish to a stronger model. Full detail in the Ling 3.0 Flash guide.
Maple Preview — the speed pick
Around 20B total with roughly 1B active at runtime, using ternary weights where every value is minus, zero or plus. That shrinks a model that would need about 38GB down to roughly 5GB. MIT licensed, 128k context, 256 experts with 8 waking per token.
Being straight: it is not good at coding — clearly behind Gemma 4, GLM 4.7 Flash and GPT-OSS in my tests. Its real story is speed and running on small devices. Full detail in the Maple Preview guide.
GLM — the heavyweight open option
The GLM line is the open-source family I’d reach for when I want more capability than a small model can give. It’s strong enough that Prime Intellect used GLM 5.2 as one of the open models in their long-context benchmarking against Claude Code and Codex.
Bigger means heavier, so this is usually an API or serious-hardware choice rather than a laptop one.
Qwen and Gemma — the dependable middle
Both are widely used open-weight families with a range of sizes, which is exactly what you want when you’re matching a model to your hardware. Gemma 4 is a solid general local pick; the Qwen line spans from small to very large.
The trap with both is reaching for the biggest size your machine can technically load. In an agent loop that makes many calls per task, an oversized model is unusable.
🔥 Want the setup already done for you? The full model stack — open-source, local and free engines with swap-in-swap-out switching — is pre-wired inside the Agent OS in the AI Profit Boardroom — install file, video tutorials, daily updates and four live coaching calls a week where you can share your screen and get it working. → Get access here
Match the model to the job, not the leaderboard
| What you’re doing | What to run |
|---|---|
| Tool calls and agent loops | LFM2.5-2.6B |
| General building and functionality | Ling 3.0 Flash |
| Fast local generation, small devices | Maple Preview |
| Heavier reasoning on open weights | GLM |
| General local use on decent hardware | Gemma 4 or a mid-size Qwen |
| Anything you’d bill a client for | A frontier model — be honest about the gap |
The single most common mistake is loading the largest model your hardware can hold. Hermes makes many calls per task, so a model that takes a minute per step is worse than useless. Start small and go up only if you need to.
Why open source matters beyond the price
- You can self-host. Nothing leaves your machine, which matters the moment client data is involved.
- Nobody can deprecate it. A closed model can be retired or changed underneath you. Open weights you’ve downloaded stay exactly as they are.
- You can inspect and fine-tune it. Not everyone will, but the option is real.
- No rate limits of your own making. Self-hosted means the only ceiling is your hardware.
- Licences vary. Maple is MIT, which allows commercial use. Always check the licence before you build a business on one.
How to run them with Hermes
- Pick one model to start — LFM2.5-2.6B if you want the safest agent-work answer.
- Install a local runtime (LM Studio is easiest) or point at a hosted endpoint.
- Create a Hermes profile for that model — see the local model setup guide.
- Test tool calls, not chat. Plenty of models converse well and fail the moment they must call a tool.
- Put a router in front if you want automatic failover — see OmniRoute or 9Router.
- Keep a second profile on a frontier model for the jobs that genuinely need it.
A note on hosted open-source models. Running an open model through somebody else’s free API is not the same as running it yourself. The weights being open doesn’t stop the host logging your prompts. If the work is confidential, self-host it or use a provider whose terms you’ve actually read.
Want the whole stack ready-made? The Agent OS in the AI Profit Boardroom ships with the local engine, model switching, the free routers and the memory system already wired together. Start free with the free AI course and community or grab 200+ AI SEO prompts.
FAQ
What are the best open source models for Hermes Agent?
LFM2.5-2.6B for agent work since it was trained through the Hermes harness, Ling 3.0 Flash as the best all-rounder, Maple Preview for speed, and the GLM family when you need more capability.
Is open source the same as free?
No. Open source means the weights are published; free means it costs nothing right now. Plenty of free models are closed, and some open models cost money to run.
Which one should I start with?
LFM2.5-2.6B. It runs on about 8GB, it’s fast, and it was specifically post-trained through the Hermes agent harness.
Can I use these for coding?
Cautiously. Ling handles functional builds well. LFM and Maple are both weak at coding — I wouldn’t use either for it.
Do I need a powerful machine?
Not for the small ones. Something in the 2–3B class runs on around 8GB. Larger open models like GLM want an API or serious hardware.
Are open source models private?
Only if you run them yourself. Using an open model through someone else’s API means their logging terms still apply.
What about licences?
They vary, and it matters if you’re building a business on one. Maple Preview is MIT, which permits commercial use. Check before you commit.
Can I run several at once?
Yes, and it’s the setup I’d recommend — one model for agent work, another for building, and a frontier model for the hard calls.
The bottom line
The best open source models for Hermes Agent are LFM2.5-2.6B for agent work, Ling 3.0 Flash as the all-rounder and Maple Preview when speed matters most — with GLM waiting when you need real weight behind it. Pick by job rather than by leaderboard, start smaller than you think, and keep a frontier model aside for anything you’d put your name on.
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.