Hermes Agent runs on whatever model you point it at, and several of those cost nothing. Here are the free routes that actually work with Hermes in 2026, what each is good for, and where the catch is.
Updated 1 October 2026: an earlier version of this post put OpenCode Free first. That route no longer works with Hermes, so it is now covered in a note at the end instead of being recommended.
1. Ollama or llama.cpp — Free Forever, On Your Hardware
Running the model yourself is the only option with no quota, no rate limit and no third party seeing your prompts. Hermes treats both as a custom OpenAI-compatible endpoint, and declares convenient aliases — ollama, llamacpp, llama.cpp, vllm and local all resolve to it.
# config.yaml
model:
provider: llamacpp
default: qwen2.5-7b-instruct
base_url: http://localhost:8080/v1A 7B model quantized to Q4_K_M needs roughly 6 GB of RAM and runs on a laptop CPU or Apple Silicon. The tradeoff is real: small local models are noticeably weaker at multi-step tool calling than a frontier model, and agent work leans on exactly that.
Best for: privacy-sensitive work, offline use, and high-volume tasks where per-token pricing would hurt.
2. OpenRouter Free Models
OpenRouter aggregates dozens of providers behind one API, and a rotating set of models carry a :free suffix. Hermes has a first-class openrouter provider, so this is a config change rather than an integration.
The free pool changes regularly and free requests are rate-limited and deprioritized against paid traffic. Treat any specific free model as temporary — check the current list rather than pinning one and expecting it to last.
Best for: sampling many models cheaply before committing to one.
3. Google Gemini Free Tier
Google offers a free tier on the Gemini API, and Hermes ships a gemini provider. You need a Google account and an API key, but no card for the free tier. Quotas and eligible models change often enough that the only reliable source is Google's current pricing page.
Best for: multimodal work — image understanding in particular — where local models struggle.
4. Hugging Face and NVIDIA NIM
Hermes ships providers for both huggingface and nvidia. Each offers a free allowance for open-weight models, both need a free account and an API key, and both are metered. They are worth knowing about mainly because they widen the range of open models you can try without running the weights yourself.
Best for: testing a specific open-weight model you cannot fit locally.
5. A Free Trial on Managed Hosting
If the point is to see whether an always-on agent is useful before investing a weekend in setup, OpenClaw Launch runs Hermes Agent with free models included and a 30-minute free trial, no card required. No model configuration, no server, and the messaging channels are already wired up.
Comparison
| Option | Account needed | API key | Real limit |
|---|---|---|---|
| Ollama / llama.cpp | No | No | Your hardware |
| OpenRouter free | Yes | Yes | Rotating pool, deprioritized |
| Gemini free tier | Yes | Yes | Daily quota |
| Hugging Face / NVIDIA | Yes | Yes | Metered allowance |
| Managed trial | Yes | No | 30 minutes |
Which Should You Pick?
- Just trying Hermes today: the managed trial, which needs no API key, or Ollama if you would rather not sign up for anything.
- Privacy or offline: llama.cpp or Ollama locally.
- Want to compare models: OpenRouter's free pool.
- Images or documents: Gemini's free tier.
- Want it running 24/7 without a server: a managed trial.
What Happened to OpenCode Free?
Hermes ships an opencode-free provider that talks to OpenCode's Zen relay without a key, and for a while it was the easiest free route. As of October 2026 the relay refuses those requests: free-tier models answer with HTTP 403 and the message "OpenCode's free tier can only be used from within OpenCode", and the old default model, laguna-s-2.1-free, is no longer supported. This is a restriction on calls from outside the OpenCode app, so it affects Hermes on any machine, not one host. If your bot is set to it, switch to one of the routes above. OpenCode's paid Go plan is a separate product and is not affected.
The Honest Caveat
Free model tiers are the first thing providers cut when capacity gets tight, and an agent is a heavier consumer than a chat window — it can fire many calls per turn. Free routes are excellent for learning, prototyping and personal bots. For anything you would be annoyed to see break, budget for a paid model or hosted inference.
Get Started
Install Hermes, run hermes model, and pick a local model or an OpenRouter free model. If you would rather skip setup entirely, the 30-minute free trial gets you a working Hermes bot with no key and no card. For hosting that is free too, see free Hermes Agent hosting options.
Running OpenClaw instead? See free AI models for OpenClaw.