Playwright and Chromium are officially identified as Hermes' most memory-hungry framework feature. The official floor rises from 1 GB to at least 2 GB with browser tools, and several simultaneous pages or profiles can need considerably more headroom.
For agents that repeatedly research websites, operate web apps, fill forms or keep several browser sessions active.Hermes Agent Hardware Requirements: CPU, RAM, Storage & Local AI
Learn Hermes Agent hardware requirements for cloud models, browser automation, Docker, subagents, 64K context and local AI on Zima hardware.
Hermes Agent hardware requirements at a glance
Size Hermes in two layers: the agent framework runs the gateway, tools, memory, skills and browser workload, while model inference can run through a hosted provider or on separate local hardware. Local inference is usually the much larger hardware requirement.
- CPU
- Hermes' official Docker guidance lists 1 CPU core as the minimum and 2 cores as recommended. Browser rendering, parallel subagents, file processing and other tool workloads benefit from additional CPU headroom.
- RAM
- Official Hermes Docker guidance lists 1 GB RAM minimum and 2โ4 GB recommended. With Playwright/Chromium browser tools active, allocate at least 2 GB. These figures cover the Hermes container, not the memory needed to serve a local LLM.
- Storage
- Hermes officially lists 500 MB as the minimum data-volume capacity and 2 GB or more recommended because sessions and skills grow over time. For an always-on ZimaOS deployment, use more SSD capacity for persistent memory, browser data, logs, workspaces and tool output.
- Runtime and deployment
- Hermes supports Linux, macOS, Windows, WSL2 and other documented targets. Upstream native installers provision the required Python, Node.js, ripgrep and ffmpeg automatically, while the official Docker image carries its own runtime stack.
- GPU and local models
- A GPU is not required when Hermes uses Nous Portal, OpenRouter, OpenAI, Anthropic or another hosted provider. Local inference is separate: Hermes requires the selected model endpoint to provide at least 64,000 tokens of context, so model weights plus KV-cache memory must both fit.
- Best Zima starting point
- ZimaBoard 2 832 is already well above Hermes' official container minimums and is a strong always-on hosted-model gateway. Choose 1664 for more browser, subagent and co-hosted-service headroom, and move to ZimaCube 2 when broader NAS, multi-container or local-AI workloads justify it.
From official requirements to the right setup
Hermes sizing starts with hosted versus local inference, then adds browser automation, parallel subagents, Docker terminal backends, persistent memory, messaging gateways and the other services sharing ZimaOS.
-
Official requirements
Decide where inference runs. With a hosted provider, Zima hardware mainly runs the Hermes gateway, tools, skills, memory and channels. With a local model, the inference server becomes a separate memory and accelerator workload.
-
Confirm your needs
Use the official Docker resource table as the framework baseline: 1 GB RAM, 1 CPU core and 500 MB data volume minimum; 2โ4 GB RAM, 2 cores and 2 GB or more data volume recommended. Browser tools raise the memory floor to at least 2 GB.
-
Leave room to grow
List local workload multipliers: Playwright/Chromium, multiple profiles or subagents, Docker sandboxes, cron jobs, code execution, document processing, voice features and other tools can consume much more CPU, RAM and disk than a chat-only gateway.
-
Run it on ZimaOS
Install Hermes from the ZimaOS App Store, configure the model provider and gateway, then monitor CPU, memory, persistent-data growth and browser/subagent concurrency during real workflows before adding local inference or exposing additional services.
Check every playback client
- Hosted model API or local model inference
- Number of profiles, subagents and messaging channels
- Browser automation and simultaneous Chromium sessions
- Local or Docker terminal backend and sandboxing needs
- Cron jobs, code execution and file-processing workloads
- Persistent sessions, memories, skills, logs and workspace growth
- Local-model context length, model size and inference backend
- Other ZimaOS apps, VMs, storage and AI services running concurrently
Official minimum requirements
Hermes publishes an explicit Docker resource table for the agent framework. These values apply to the Hermes container and persistent data volume; they must not be confused with the much larger hardware requirements of a locally hosted LLM.
Use 1 GB RAM, 1 CPU core and 500 MB data volume as the current official container minimums, with 2โ4 GB RAM, 2 cores and 2 GB or more storage recommended. Allocate at least 2 GB RAM when browser tools are enabled. If the model runs locally, size inference separately around Hermes' mandatory 64K context floor.
| Requirement | Official minimum | What this supports |
|---|---|---|
| Container memory | 1 GB minimum; 2โ4 GB recommended | These are official Hermes Docker framework figures. They do not include the memory required by a local model server running on the same system. |
| CPU | 1 core minimum; 2 cores recommended | Hermes is relatively light in hosted-model mode, but browser rendering, parallel work and CPU-heavy tools can justify additional cores. |
| Persistent data volume | 500 MB minimum; 2 GB or more recommended | The data directory stores configuration, sessions, memories, skills, cron jobs, logs and per-profile state, so a long-running installation can grow well beyond the minimum. |
| Browser automation | At least 2 GB RAM with Playwright/Chromium active | Hermes calls browser automation its most memory-hungry framework feature. More concurrent pages or profiles can justify additional headroom beyond the official floor. |
| Hosted model inference | No local GPU required | Hermes can use Nous Portal, OpenRouter and other hosted or custom providers. In this architecture the local system runs the agent framework while the provider performs model inference. |
| Local model context | At least 64,000 tokens required for agent use with tools | Hermes rejects smaller model context windows at startup. Local-model sizing therefore has to include model weights plus KV-cache memory at 64K or more, not just the model file size. |
When to upgrade your hardware
Upgrade Hermes hardware when a defined workload multiplier appears. Browser automation, parallel agents and local inference are much better upgrade signals than simply installing more skills.
Browser automation becomes a routine workload
Profiles, subagents and tool jobs run in parallel
You move the LLM onto local hardware
Hermes supports multiple profiles, isolated subagents and several execution backends. Parallel browser, terminal and code workloads increase process count, RAM use and disk activity even when all LLM inference remains in the cloud.
For multi-channel assistants, scheduled workflows, code agents and households running several persistent agent profiles.This is the largest hardware jump. Hermes requires at least 64K model context, so local inference must hold both the model weights and a large KV cache. A model that technically loads with a short chat context may still be unsuitable for Hermes agent workflows.
For users moving from hosted APIs to Ollama, llama.cpp, vLLM, LM Studio or another local OpenAI-compatible model server.Plan hardware growth with confidence
Hermes scales most cleanly when the gateway, persistent state, browser/sandbox workload and model inference are treated as separate resources.
Keep hosted inference separate from the always-on gateway
Hermes' official repository explicitly supports running on low-cost VPS-style infrastructure because model calls can be remote. This keeps the Zima system focused on uptime, tools, messaging and persistent memory instead of GPU inference.
Use ZimaBoard 2 for the gateway and increase CPU/RAM only when browser, subagent or other ZimaOS workloads create real pressure.Give persistent Hermes data more than the minimum disk
The official 500 MB minimum only describes the initial data-volume floor. Sessions, memories, skills, logs, browser state, downloaded files and agent-created artifacts accumulate during long-running use.
Use SATA or NVMe SSD storage for an always-on Hermes deployment rather than treating the onboard system storage as unlimited workspace.Run local models on a separate inference endpoint when practical
Hermes can connect to OpenAI-compatible local endpoints, including Ollama, llama.cpp and vLLM-style services. Separating the model server lets a small always-on gateway remain responsive while a larger GPU system handles model weights and 64K-plus context.
Keep Hermes on ZimaBoard 2 or ZimaCube 2 and connect it over the LAN to a dedicated local inference system when the model exceeds the Zima device's practical memory or GPU budget.Preserve isolation and authentication as the agent grows
Hermes can execute terminal tools, access files, hold credentials and expose a dashboard/API. Official Docker guidance now requires authentication for non-loopback dashboard exposure and supports Docker isolation for tool execution.
Do not remove sandboxing or security controls simply to fit an undersized host; reserve enough CPU, RAM and storage for the execution boundary the workflow requires.Can it run on ZimaOS?
Hermes is currently available in the ZimaOS App Store, and ZimaSpace publishes a dedicated Hermes setup guide. The ZimaOS package runs Hermes in a container, which makes the official Docker framework requirements the most relevant upstream hardware baseline.
Install Hermes from the ZimaOS App Store
ZimaOS currently lists Hermes as an AI application. Install it from the App Store, then use the containerized Hermes environment to configure the model provider, tools and messaging gateway.
Open Hermes in the ZimaOS App StoreConfigure Hermes inside its ZimaOS container
ZimaSpace's current setup guide enters the Hermes container as the hermes user, activates the packaged virtual environment and runs hermes setup. This confirms that the ZimaOS deployment should be treated as a containerized gateway rather than a host-level manual Python install.
Read the ZimaOS Hermes setup guideKeep local-model sizing separate from the ZimaOS app
The App Store package can point Hermes at a hosted or custom model provider. If the custom endpoint is a local LLM server, validate its 64K context, model memory and inference performance separately from the Hermes container requirements.
Read Hermes model-provider guidanceChoose Zima hardware for your Hermes Agent workload
The first decision is whether Zima only runs the Hermes framework or also carries substantial browser, subagent, storage and local-AI workloads. Hosted-model Hermes is light enough that the smallest current ZimaBoard 2 already has large headroom over the official container baseline.
Will primary model inference stay on a hosted provider or separate inference server?
Yes โ Zima runs Hermes gateway, memory, tools and browsers
No local GPU is required. ZimaBoard 2 832 already provides much more CPU and RAM than the official Hermes container baseline; move to 1664 when browser/subagent concurrency or other ZimaOS services need additional memory.
- Hosted-model Hermes with normal browser, tool and messaging useZimaBoard 2 832
- More simultaneous browsers, profiles, subagents and co-hosted appsZimaBoard 2 1664
No โ heavier home-server or local-AI workloads share the system
Use ZimaCube 2 when Hermes is part of a broader multi-drive, multi-container or local-AI server. Creator Pack provides a dedicated GPU path, but local-model suitability still has to be validated against the exact model, GPU memory and Hermes' 64K context floor.
- Hosted-model Hermes plus larger NAS and heavier CPU-side servicesZimaCube 2 Standard
- Higher multi-agent/container concurrency with more CPU and memory headroomZimaCube 2 Pro
- Local-model experimentation plus genuine GPU, AI or VM workloadsZimaCube 2 Creator Pack
This is a workload guide, not a guaranteed agent-count, browser-session or tokens-per-second benchmark. Results depend on Hermes version, model provider, browser pages, profiles, tool workloads, sandbox backend, local-model size and context, quantization and other ZimaOS services.
| Zima hardware | Best for | Example workload | Core configuration | Recommended boundary | Next step |
|---|---|---|---|---|---|
| ZimaBoard 2 832 | An always-on Hermes Agent gateway using hosted model APIs, browser tools, messaging and ordinary persistent memory workloads. | One or a few personal agent profiles, scheduled jobs, messaging channels, light-to-moderate browser automation and normal ZimaOS services. |
|
Eight gigabytes can become constrained only when browser sessions, parallel subagents, Docker workloads or other ZimaOS apps consume substantial memory. Do not treat the 32 GB eMMC as unlimited long-term agent workspace. | Get Now |
| ZimaBoard 2 1664 | A hosted-model Hermes server with more browser automation, profiles, subagents and co-hosted ZimaOS services. | Always-on multi-channel agents, parallel tool workflows, Docker-backed execution and a broader compact homelab. |
|
Sixteen gigabytes is useful for framework-side concurrency and other containers but does not make the N150 a high-performance local LLM platform. Local inference still needs separate model/context sizing. | Get Now |
| ZimaCube 2 Standard | Hermes hosted-model workloads combined with a larger NAS, persistent datasets, backups and heavier CPU-side home-server services. | Agent gateway, browser/tool automation, local file datasets, backups and multiple ZimaOS applications where six-bay storage matters. |
|
Standard has the same 8 GB system-memory class as ZimaBoard 2 832. Choose it for stronger CPU and integrated multi-drive storage, not because Hermes itself needs a six-bay NAS. | Get Now |
| ZimaCube 2 Pro | Higher Hermes gateway-side concurrency combined with many containers, large local datasets and broader NAS/home-server workloads. | Several profiles or subagents, browser automation, Docker execution, scheduled workflows, file processing and other always-on ZimaOS services. |
|
More CPU and 10GbE do not replace GPU or unified-memory capacity when model inference runs locally. The mandatory 64K context can make local-model memory much larger than the model file alone. | Get Now |
| ZimaCube 2 Creator Pack | Hermes combined with local-model experimentation, GPU-accelerated AI services, creator applications or heavy virtual machines. | Agent gateway plus a local model endpoint or other GPU workloads where 64 GB system memory and a dedicated NVIDIA GPU are independently useful. |
|
Do not present Creator Pack as a universal high-end Hermes local-model solution. Large agentic models and long contexts can exceed the GPU memory of a single workstation GPU and may need a separate higher-end inference system. | Get Now |
What the Press Says
Highlights from trusted reviewers worldwide.
โZimaCube 2: Not just another NAS, tested with 25TB storage, local AI agents, 4K transcoding, and real homelab workflows.โRead full review
โThe ZimaBoard 2 is a compact x86 server board that can be turned into a mini NAS, home server, media box, or self-hosting hub.โRead full review
โZimaCube 2: A modern, high-performance NAS with plenty of room to growโbuilt for users who want more than basic storage.โRead full review
โCoverage focused on ZimaCube 2's open hardware foundation, no monthly fee, and self-hosting flexibility.โRead full review
Loved by the Community
Stories and reviews from people who build with Zima every day.
โ
โ
โ
โ
โ
Zima Blade Little yet Powerful
Maybe I am not digital natives but I live with PCs since 12 years old in 1984 when IBM PC clone come to my home. Many years have passed and many operating system I've tried. For me Zima blade and CasaOS was a quantum leap for home PC enthusiast and server lab machine to make me stay curious and relevant for this era.
โ
โ
โ
โ
โ
Very good!!
I use ZimaCube Pro as 5th Proxmox cluster node. It runs several VMs and containers, including a VM with GPU passthrough to run a self-hosted LLM. A specific LXC container runs a Samba server for NAS capabilities using four of six RAID 6 SATA HDDs with ZFS.
โ
โ
โ
โ
โ
Great innovation for mini server!
It is very useful and makes a powerful mini server for many purposes, including university and college students in engineering and electronics. Thank you so much for making this server.
โ
โ
โ
โ
โ
Avaliaรงรฃo ZimaBoard 2
Construรญ um servidor de uso pessoal. O desempenho estรก muito bom e funciona perfeitamente onde quer que eu esteja. A surpresa รฉ nรฃo dependermos de grandes estruturas para termos nosso prรณprio servidor de dados. Como iniciante, estou gostando bastante do ZimaOS, pois ele รฉ simples e eficiente.
Frequently asked questions
These answers separate Hermes framework requirements from browser, multi-agent and local-inference workloads so model hardware is not confused with the lightweight gateway itself.
How much RAM does Hermes Agent need?
Official Hermes Docker guidance lists 1 GB RAM minimum and 2โ4 GB recommended for the framework. Browser automation raises the official allocation floor to at least 2 GB. A local LLM needs its own additional memory budget and should be sized separately.
Is 8 GB RAM enough for Hermes Agent?
Yes for many hosted-model home deployments. Eight gigabytes is well above Hermes' official 2โ4 GB recommended container range and leaves room for normal browser, messaging and tool workloads. Heavy parallel browsers, subagents and other ZimaOS services can justify 16 GB.
Does Hermes Agent need a GPU?
No when Hermes uses a hosted provider. The GPU only becomes a requirement when the same machine also serves a local model or another GPU-accelerated workload. This framework-versus-model distinction is the most important hardware decision.
Can ZimaBoard 2 run Hermes Agent?
Yes. Hermes is currently available in the ZimaOS App Store, and ZimaBoard 2 832 provides 4 CPU cores and 8 GB RAMโwell above Hermes' current official container minimum and recommended figures. The 1664 adds memory for heavier browser and multi-service use.
How much RAM does Hermes browser automation need?
Official Hermes Docker documentation says Playwright/Chromium is the most memory-hungry framework feature and requires at least 2 GB RAM when active. The exact requirement rises with page complexity, concurrent browser sessions and other profiles or tools.
Can ZimaBoard 2 run a local model for Hermes?
It can run some small CPU-based or heavily quantized models, but that is not equivalent to a strong Hermes agent experience. Hermes requires at least a 64K context window, and CPU-only generation with enough model and KV-cache memory can be slow. Hosted inference or a separate GPU model server is usually the better ZimaBoard 2 architecture.
Why does Hermes require at least 64K context for local models?
Hermes' official provider documentation says models below 64,000 tokens are rejected because the system prompt, tool schemas and working conversation need enough room for reliable multi-step tool use. For local inference, that means memory must cover both model weights and the KV cache for at least 64K tokens.
When should I choose ZimaCube 2 instead of ZimaBoard 2 for Hermes?
Choose ZimaBoard 2 for an efficient hosted-model gateway and normal browser/tool automation. Choose ZimaCube 2 when Hermes shares the system with multi-drive storage, many containers, larger datasets or heavier concurrency. Creator Pack is only justified when a real GPU/local-AI workload is also part of the plan.
What sources and further reading informed this Hermes Agent hardware guide?
The six references below provide useful 2026 perspectives but are not equally authoritative. OpenClaw Launch gives a practical 1 GB / 2โ4 GB framework tier, but some software details such as Node 20 and WSL2-only Windows guidance lag current upstream Hermes documentation. OpenAI Tools Hub closely tracks the current official Docker resource table and correctly highlights Hermes' mandatory 64K local-model context, though its per-model memory calculations are derived rather than official Hermes benchmarks. BIZON focuses on much larger local-AI workstation designs and should not be used as the minimum for hosted-model Hermes. Julian Goldie correctly emphasizes the cloud-versus-local split but provides few numeric framework limits. OpenClawDC's 1โ2 GB harness tier is a practical interpretation rather than the canonical specification. hermes-agent.org provides a quick-start mirror/landing page, but where it differs from the current Nous Research documentationโfor example native Windows supportโthe canonical Nous Research docs take precedence.
