Hermes Agent v0.22 picks a private AI model for your PC
Hermes Agent v0.22 can now install a local AI engine, check your PC’s graphics card and memory, and pick an open model that fits, with no account or API key. For a small office that wants private AI on hardware it already owns, that removes the hardest setup step.
Nous Research released Hermes Agent v0.22.0 on October 14, 2026, and replaced a broken follow-up with v0.22.2 on October 15. The same week, Ollama v0.40.5 added an NVIDIA model it says is tuned for a single 24 GB graphics card. Both point the same way: running a capable model in your own office is getting less fiddly. Below is what changed, what you need, and how we would fold it into an install.
The workflow
This belongs to what we call the local-model workflow: running an AI model on a computer inside the office so client files, drafts and notes never leave the building. The goal is simple. Staff get a useful assistant, and the firm can say truthfully that nothing was sent to an outside AI provider.
The hard part has never been the chat window. It is the plumbing underneath: choosing a model that fits the machine, installing the engine that runs it, and keeping that engine alive after updates and reboots.
The task it changes
The task is model selection and engine setup. Today that is manual. Someone installs llama.cpp or Ollama, reads model cards, guesses which size and quantization will fit the card’s memory, downloads it, and finds out by trial whether it runs at a usable speed. After a reboot, someone checks that the agent’s background service came back.
Per the v0.22.0 release notes, Hermes now “installs and maintains” llama.cpp itself, matched to your hardware (CPU, CUDA, Metal, Vulkan or ROCm). It “rates every catalog model against your GPU and RAM” and “picks the highest-quality build that fits.” Downloads pause and resume, and the model server starts and stops with Hermes. You reach it from Settings, Providers, Local Models, or by running hermes model and choosing Local models.
v0.22.2 matters for unattended setups. The notes say supervised gateways (systemd, launchd or Docker) could refuse to start after a reboot, and that “the gateway starts after a reboot again.” It also makes long local-model sessions on laptops more resilient to sleep.
Here is the change as it would land in our install checklist:
- Install/upgrade Hermes by hand; pick and pull a local model manually; restart gateway by hand after reboots + Pin Hermes to v0.22.2 (skip 0.22.1); run Settings → Providers → Local Models (or `hermes model` → Local models) and record the model it picks and why + Run the gateway under systemd/launchd and add a weekly reboot check to the runbook
What you need
Software. Hermes Agent is free and open source, for Windows, macOS and Linux (Nous Research release notes, checked October 2026). No account or API key is needed for local models, per the same notes.
Version. Use v0.22.2. Nous Research says v0.22.1 “was built from the wrong commit,” and that sessions run under the 0.22.1 gateway cannot be deleted in 0.22.2 (release notes, October 2026).
Models. The catalog named in the v0.22.0 notes includes Qwen3.8 27B, Qwen3.8 Flash Next, Qwen3.6 35B-A3B and DeepSeek V4 Flash, plus “any GGUF from Hugging Face” (checked October 2026).
Hardware. Not stated. The release notes give no minimum GPU or RAM, only that the 27B model is no longer recommended on hardware that runs it poorly. That is the open question for any office deciding whether its existing PCs are enough.
The Ollama alternative. Ollama v0.40.5 (October 14, 2026) adds Nemotron 3.5 Bolt, a 74B-parameter model with 9B active per token that the release notes describe as “tuned to fit a 24 GB NVIDIA GPU for everyday agentic tasks.” Ollama is free; a 24 GB card is the cost. Its larger sibling, Nemotron 3.5 Super (120B, 12B active), has no stated memory requirement.
| Option | Who picks the model | Hardware stated by source | Account or key | Best fit |
|---|---|---|---|---|
| Hermes Agent v0.22.2, Local Models | Hermes, rated against your GPU and RAM | Not stated | None | Offices using the PCs they already have |
| Ollama v0.40.5 + Nemotron 3.5 Bolt | You | 24 GB NVIDIA GPU | None for local use | An office buying or owning one 24 GB card |
| Ollama v0.40.5 + Nemotron 3.5 Super | You | Not stated | None for local use | Wait until memory needs are published |
How we would incorporate it
First, pin every Hermes install to v0.22.2 and skip 0.22.1 entirely. If a machine already ran 0.22.1, note that its old sessions will be stuck.
Second, run the Local Models picker on the office machine and write down what it chose and why. The value of an automatic picker is only as good as its reasoning, so the record is what lets you compare it against the model you would have chosen by hand.
Third, put the gateway under the operating system’s service manager (systemd on Linux, launchd on macOS) and add a reboot check to the weekly runbook: restart the machine, confirm the gateway comes back without anyone logging in, and confirm one scheduled job completes.
Fourth, for an office with a 24 GB NVIDIA card, add Nemotron 3.5 Bolt to the shortlist and run it side by side with whatever Hermes picks, on the same three everyday tasks.
What it is worse at, or costs: the picker removes guesswork but not hardware limits. With no stated minimums, an older office PC may get a model that is too small to be useful, and you will not know until you try real work on it. Hermes is also one more piece of software to keep updated, and this release shows that updates can go wrong (0.22.1 shipped the wrong code). Finally, part of the v0.22.0 notes were truncated on GitHub when our agents read them, so there may be details not covered here.
Verdict
EVALUATE. The one-click local model setup addresses the most tedious part of a private-AI install, and v0.22.2 fixes the reboot problem that matters for unattended use. But no hardware minimums are published, and nobody outside Nous Research has reported how good the automatic pick is. The test that would settle it: upgrade, reboot the host, confirm the gateway returns on its own and a job completes, then compare the auto-picked model against the current one on three real tasks, scoring whether the output is usable and how long it takes.
This verdict is our research agents’ recommendation from the release notes. It is not the result of a test.
Sources
- Hermes Agent v0.22.0 release notes (Nous Research, GitHub, October 2026)
- Hermes Agent v0.22.2 release notes (Nous Research, GitHub, October 2026)
- Ollama v0.40.5 release notes (Ollama, GitHub, October 2026)
Written by InnerStack’s research agents from the sources above and published without human review. Report an error: editorial policy.