GoldieBench Blog · 10 min read
Hermes Agent v0.22: What's New In The Threshold Release, How To Get It, And The Models To Run In It
Hermes Agent v0.22 explained: what's new in the Threshold Release, how to update, every headline feature, and how the models you can run in it score.
Hermes Agent v0.22 is the newest major release of the free, open-source AI agent from Nous Research, and Nous calls it "The Threshold Release".
The official release notes are dated 13 October 2026, and a v0.22.1 patch followed on 14 October 2026.
A fresh desktop install now sets itself up through a chat on a free trial model, Hermes can work inside apps such as Linear and Notion, and local models install in one click.
It also brings a 600-plugin catalog, password-manager sign-in for the browser, a bot screen you can take over and a new wave of Claude 5.5 and GPT-6 models.
I'm Julian Goldie, and the video above is my walkthrough of the release notes rather than a hands-on test of every feature.
In this post I'll explain what v0.22 is, how to get it and how each feature works, and then I'll show you how the models you can run in it score on GoldieBench.
What is Hermes Agent v0.22?
Hermes Agent is an AI agent that runs on your own computer or server, uses tools, learns skills and works through chat apps.
Version 0.22 is the biggest window Hermes has shipped, and Nous says it is about 2.5 times the size of the v0.21.0 cycle.
These are the facts from the official notes.
| Item | Detail |
|---|---|
| Release | Hermes Agent v0.22.0, "The Threshold Release". |
| Maker | Nous Research builds and maintains Hermes Agent. |
| Date | The release notes are dated 13 October 2026. |
| Patch | v0.22.1 followed on 14 October 2026. |
| Price and licence | It is free and open source under the MIT licence. |
| Work since v0.21.0 | The notes count about 24,200 commits, 6,080 merged PRs, 7,180 closed issues and 2,520 contributors. |
| Codebase | The source shrank from about 1.06 million lines to about 698,000, a cut of 34%. |
| Roll-up | It also documents the v0.21.1 to v0.21.7 patch releases. |
In my video I grouped it into three moves, which were walk in, plug in and hand off.
Walk in means getting started, plug in means connecting your apps and tools, and hand off means letting the agent carry on while you step in only when needed.
How to get Hermes Agent v0.22
The desktop app is now a sealed, signed package that updates itself, so you accept the update when it appears.
Microsoft Store installs update through the Store.
A command-line install updates with hermes update.
A fresh install on Linux, macOS or WSL2 uses curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash.
A fresh install on native Windows uses iex (irm https://hermes-agent.nousresearch.com/install.ps1) in PowerShell.
Docker users pull nousresearch/hermes-agent:stable.
Source installs now follow tested SemVer releases by default, and hermes update --set-channel main moves you back to every commit if you want the bleeding edge.
The upgrade notes to read first
Managed installs now run on Python 3.14, and the package manager installs it for you.
The Windows desktop app is MSIX-only and needs Windows 11 22H2 or newer, and the default macOS build is Apple Silicon only.
One messaging gateway per machine is now the default.
Chats no longer reset on an idle or daily timer, so you start fresh with /new or /reset.
Ended sessions inactive for 90 days are deleted at startup unless you set sessions.auto_prune: false.
Memory providers such as Honcho, Supermemory, Hindsight and Mem0 moved into the plugin catalog, and existing users are migrated with their data intact.
Walk in: the first-launch setup chat
A fresh desktop install now opens on a short setup chat instead of a settings screen.
One card at a time, it asks for your name, an accent colour and a layout, then which apps to connect and which plugins to install.
It ends with a quick tour and up to four first tasks picked for you.
The chat runs on a free trial model called nous/welcome, once per operating-system user, so it needs no key.
The agent also gained a set_theme tool, so "make it violet, dark mode" works in any desktop chat.
The CLI, the TUI and the gateway get no free model, and existing installs do not get the setup chat automatically.
Walk in: local models in one click
Hermes can now run open models entirely on your machine, with no account and no API key.
It installs and maintains llama.cpp with the right engine for CPU, CUDA, Metal, Vulkan or ROCm.
It rates every catalog model against your GPU and RAM, and it only recommends models that fit fully in GPU memory.
You start from Settings → Providers → Local Models, or from hermes model then Local models in a terminal.
The catalog includes Qwen3.8 27B, Qwen3.8 Flash Next, Qwen3.6 35B-A3B and DeepSeek V4 Flash, plus any GGUF from Hugging Face.
Walk in: stable releases and signed packages
Every release now passes source CI, Docker, Nix, signed desktop builds and install and upgrade runs on every operating system before it ships.
Windows gets a signed MSIX for x64 and ARM64, and macOS gets a signed, notarised Apple Silicon build.
A new package manager, hermes pm, builds each update in a fresh environment and switches only once it passes a check.
That means a failed or killed update can no longer leave you half-applied.
Plug in: connectors
Hermes can now act inside hosted apps such as Linear and Notion through the Nous tool gateway.
You sign in once in your browser, and the tools are usable in the same turn.
You do not need a Nous account, because Hermes creates a free connectors-only identity the first time.
The new Connectors page lists hosted apps and your MCP servers together, and it replaces the MCP tab.
You can connect several named accounts for one app and manage them with /connectors or hermes connectors.
Plug in: the 600-plugin catalog
The plugin catalog launched this cycle and grew to 600 entries, each pinned to a reviewed commit.
You can install them from Hermes Desktop or the CLI, and browse them on the Hermes plugin docs.
The notes count 180 tools, 140 desktop panes, 65 memory plugins, 48 model providers, 34 web and search plugins, 32 platforms and 18 voice plugins.
Those 48 model-provider plugins are worth noting, because they are another way to plug a different brain into Hermes.
Plug in: password-manager sign-in
The browser agent can log into sites with 1Password, Bitwarden or Hermes's own encrypted vault.
The model never sees the password.
One-time codes are filled without entering the conversation, and card and address fills need your confirmation.
Hand off: Bot Screen, one gateway and live subagents
A bot running on a Linux gateway can now have its own desktop that Hermes Desktop streams live.
When it hits a login, a 2FA prompt or a CAPTCHA, you take the screen, finish that step and hand control back.
With a Docker, SSH or Singularity backend, the screen and the browser run inside the sandbox, and the notes say Bot Screen is off by default.
Hermes now runs one gateway and one desktop backend per machine for every profile, and each profile keeps its own credentials, cron jobs and sessions.
Running subagents stay visible with their task, elapsed time and live transcript, plus Steer and Stop buttons.
Hand off: reasoning replay and provider slowdowns
Prior reasoning is now replayed by default on every provider that can read it, across resume, branches and compaction.
When a provider has a bad hour, Hermes waits out short 503, 529 and timeout errors in visible, bounded cycles instead of ending your turn.
These two changes matter more for long jobs than any single new model does.
The smaller changes in v0.22
All user-facing text now runs through one translation catalog with full coverage in 16 bundled languages.
About 19 rarely used tools now load only when the agent searches for them, which roughly halves the desktop tool schema on every call.
Unpinned cron jobs now use your main model when they fire, and a run that never reached the model retries after 5, 15 and 30 minutes.
The model picker shows a chip such as "17% left · resets 3:00 PM" and lets you set reasoning effort as you choose a model.
What the v0.22.1 patch changed
Nous Research describes v0.22.1 as a patch with a handful of fixes for the Windows Store, plugin updates and more.
It adds a curated model list at the top of the CLI model picker.
It also adds catalog plugins, including Korean and Brazilian Portuguese language packs.
Does Hermes Agent v0.22 ship its own model?
No, it does not ship a model you would run your work on.
Hermes is the agent, so the brain is whichever model you connect, whether that is hosted or local.
The one exception is nous/welcome, a free trial model that only powers the first-launch setup chat on the desktop.
That is why the useful question for v0.22 is which model you pick, and that is where GoldieBench helps.
GoldieBench scores models on the same set of one-shot builds, with a judge rating each one from 0 to 10.
The v0.22 models already on GoldieBench
These are the models named in the v0.22 notes that I have already benched, with their live averages.
| Model | Where it appears in v0.22 | GoldieBench average |
|---|---|---|
| Claude Opus 5.5 | It is part of the Claude 5.5 wave, and it is now also on native Anthropic and Bedrock. | 7.57 across 50 tasks |
| Grok 4.7 | It is in the v0.22 model catalog additions. | 7.15 across 20 tasks |
| MiMo-V2.6 Pro | Xiaomi MiMo v2.6 is listed in the notes, and Pro is the version I benched. | 6.35 across 50 tasks |
| DeepSeek V4 Flash | It is in the one-click local model catalog. | Benched on 50 tasks and currently unranked. |
One honest surprise is worth calling out.
Claude Opus 5.5 averages 7.57, which is below Claude Opus 5 at 8.27 on the same kind of builds.
Newer does not automatically mean better at one-shot builds, so do not assume the newest name in the picker is the best builder.
The v0.22 models not benched yet
Most of the new wave is not on the board yet, and I will not guess their scores.
Claude Fable 5.1, Sonnet 5.5 and Haiku 5.5 are not benched.
GPT-6 Astra, Sol and Luna, GPT-6.1 Sol and Gemini 4 Argon are not benched.
Muse Spark 1.3, GLM-5.3-FlashX and DeepSeek V4.1 Flash are not benched either.
The closest benched relatives give you a rough guide until they are.
| New in v0.22 | Closest model on GoldieBench | Its live average |
|---|---|---|
| Claude Fable 5.1 | Claude Fable 5 | 8.10 |
| Claude Sonnet 5.5 | Claude Sonnet 5 | 7.01 |
| GPT-6 Sol and GPT-6.1 Sol | GPT-5.6 Sol | 8.16 |
| Gemini 4 Argon | Gemini 3.6 Flash | 7.08 |
| Muse Spark 1.3 | Muse Spark 1.2 | 7.55 |
| GLM-5.3-FlashX | GLM-5.2 | 7.77 |
A relative is only a hint, because a new version can score higher or lower than the one before it, as Opus 5.5 shows.
Local models in v0.22 versus my local board
The one-click local catalog leads with Qwen models and DeepSeek V4 Flash.
The Qwen 3.8 on GoldieBench averages 8.10, but that is Alibaba's hosted flagship benched through Qoder, not the 27B build Hermes runs locally.
The closest local model I have benched is Qwable 5 27B Coder, a fine-tune on a Qwen3.6-27B base, which averages 7.14.
That is still a different model from the Qwen3.6 35B-A3B in the Hermes catalog, so treat it as a rough guide only.
The local models board shows everything I have run on my own machine, and the best local model for Hermes Agent goes deeper.
I have not benched Qwen3.8 27B or Qwen3.8 Flash Next yet.
How I would choose a brain for Hermes Agent v0.22
I would start with the strongest benched model your budget allows for the main agent.
On the board today that means Claude Opus 5 at 8.27, Hermes MoA at 8.17 or GPT-5.6 Sol at 8.16.
Hermes MoA is worth a look because it is a panel of frontier models merged by a chair inside Hermes itself.
For cheap or private sub-agent work, I would use the one-click local models and keep the hard reasoning on a hosted model.
I would not pick on score alone, because one-shot builds and long agent sessions are different skills.
You can put any three models side by side on the compare page, and my best LLMs for Hermes Agent post ranks the options.
The v0.22 model picker makes this easier, because it shows when a subscription is running low before you hit the wall.
Who should update to Hermes Agent v0.22
Beginners should install it, because the setup chat is the easiest way into Hermes so far.
Existing users should update, but they should set the auto-prune option first if they keep old chats.
People on Intel Macs or Windows 10 should use the command-line install rather than the new desktop package.
Anyone running several profiles should update, because the single gateway removes a lot of process clutter.
Also On Our Network
- Agent Operatorsthe step-by-step Hermes Agent v0.22 update and setup
- AI Income Deskseven ways a business can use Hermes Agent v0.22
- AI Tool Verdictevery Hermes Agent v0.22 feature, scored with a verdict
- agentos.guidethe Hermes Agent v0.22 quick start for an Agent OS
- agentos.guidemy Hermes crash course on agentos.guide
If you run an SEO business and want to talk through where agents like Hermes fit, you can book a free SEO strategy session with my team.
Update, read the upgrade notes, then pick the brain from the live board, and you will get the most out of Hermes Agent v0.22.
FAQ
What is Hermes Agent v0.22?
Hermes Agent v0.22 is a major release of the free, open-source Hermes Agent from Nous Research, called The Threshold Release. The release notes are dated 13 October 2026. It adds a first-launch setup chat on a free trial model, connectors, one-click local models, stable release channels, signed desktop packages and a 600-plugin catalog.
How do I update to Hermes Agent v0.22?
The desktop app updates itself, the Microsoft Store version updates through the Store, a command-line install updates with hermes update, and Docker users pull nousresearch/hermes-agent:stable.
Does Hermes Agent v0.22 come with its own AI model?
Not one you would run your work on. Hermes is the agent, and the brain is whichever model you connect. The only built-in model is nous/welcome, a free trial model that powers the first-launch setup chat on a fresh desktop install.
Which v0.22 models are on GoldieBench?
Claude Opus 5.5 averages 7.57, Grok 4.7 averages 7.15 and MiMo-V2.6 Pro averages 6.35. DeepSeek V4 Flash has been benched on 50 tasks and is currently unranked. Claude Fable 5.1, the GPT-6 family and Gemini 4 Argon are not benched yet.
What is the best model to run in Hermes Agent v0.22?
On one-shot builds, Claude Opus 5 averages 8.27, Hermes MoA averages 8.17 and GPT-5.6 Sol averages 8.16. Start with the strongest model your budget allows for the main agent, and use one-click local models for cheap or private sub-agent work.
What changed in Hermes Agent v0.22.1?
v0.22.1 is a patch from 14 October 2026 with Windows Store fixes, a curated model list at the top of the model picker and new catalog plugins, including Korean and Brazilian Portuguese language packs.
Keep reading
Hermes Herald AI Agent: Which Brain To Run It On, With Real GoldieBench Scores
Herald OS: What It Is, How To Set It Up, And Which Benchmarked Models To Run Under It
AI SEO Optimization Agency: What The Work Is, Who To Hire, And How The Models Behind The Answers Score🌐 Sister-site take: read this on agentos.guide