You want to run your own AI chatbot, but the system requirements leave you baffled—some say 256MB of RAM is enough, others insist on at least 8GB, and some claim 16GB is the starting point. Why is there such a huge discrepancy?

It all comes down to what you are running and how you are running it. The system requirements for Hermes Agent aren't a fixed number; they depend on the operating mode you choose. Once you understand this, you can build a capable machine within a reasonable budget.

First, let's clarify: what exactly is Hermes Agent running?

Hermes Agent is a self-evolving AI agent framework developed by Nous Research. Its core capabilities include persistent memory across sessions, the automatic generation of reusable skills, and parallel tool invocation. Its fundamental difference from ordinary chatbots is that it learns from every task—becoming smarter the more you use it.

However, there is a crucial distinction here: Hermes Agent itself does not include a Large Language Model (LLM). It acts as an "orchestration framework" responsible for managing conversations, invoking tools, maintaining memory, and executing tasks. The "brain" that actually generates responses requires you to provide a model—this could be a cloud-based API (like Kimi, Claude, or GPT) or a locally hosted model (via Ollama).

This distinction directly determines the hardware specifications you need.

Two operating modes, with system requirements differing by an order of magnitude

Mode 1: Framework Mode (Invoking Cloud APIs)

This is the mode most users should choose. Hermes Agent handles only orchestration and memory management, while model inference is offloaded to a cloud API.

The official minimum requirements are 256MB of RAM and 100MB of storage. While this figure might seem incredible, the logic is sound: the framework itself is simply a Python program that doesn't need to load any model weights; memory usage is primarily driven by managing conversation context and reading/writing to the SQLite database.

Recommended configuration: 2-core CPU, 2GB RAM, 1GB storage, and a network speed of 10Mbps or higher. 2GB of RAM is a reasonable baseline for smooth operation—while 1GB might struggle with concurrent sessions, long conversation contexts, or simultaneous tool calls, 2GB provides plenty of headroom.

Official recommendations for Docker deployment further confirm this range: 1GB of RAM is the minimum requirement for the container, while 2–4GB is the recommended configuration. If browser tools are enabled, it is recommended to allocate at least 2GB of RAM.

The cost of this mode is very low. A VPS with 2 CPU cores and 2GB of RAM typically costs between $30 and $50 per year—less than $4 per month. The official Hermes Agent documentation explicitly states that it "can run on a $5 VPS."

Mode 2: Local Model Mode (Ollama + Local Inference)

If you want to keep your data entirely on your own server or avoid API call fees, you can run models locally using Ollama. However, hardware requirements increase sharply in this scenario.

RAM requirements depend on the model size. Running a small model with 3 billion parameters (3B) requires a minimum of 8GB of RAM and at least 5GB of storage. For larger models (27B+ parameters), 32GB+ of RAM and 30GB+ of storage are recommended. The RAM required to load a model is roughly 1.2 times the model file size; for instance, an 8GB model file actually requires nearly 10GB of available RAM to run.

For local model mode, 16GB of RAM is the baseline for "usability," while 32GB is the threshold for a "good experience." An 8-core CPU is also recommended, as the computational load for local inference falls entirely on the CPU (unless GPU acceleration is available).

The cost of this mode is significantly higher. A VPS with 16GB of RAM typically costs over $40–$60 per month. Yet, the performance might still be inferior to calling cloud APIs; running a 7B model on a 16GB VPS results in noticeably slower response times compared to using APIs like Kimi or Claude.

My recommendation: Choose the framework mode unless you have a specific requirement otherwise.

For the goal of running a local AI chatbot, the framework mode (using cloud APIs) offers the best balance of cost-effectiveness and user experience.

There are three reasons for this. First, there is a massive gap in model quality. Cloud-based models like Kimi and Claude far outperform any small model you could run on a VPS; they are in a different league regarding conversation quality, tool-use accuracy, and long-context understanding. Second, the cost is lower. A 2-core, 2GB VPS costs less than $50 per year, whereas a VPS capable of running a 27B model costs tens of dollars per month. Third, maintenance is simpler. Running a local model involves handling a series of engineering challenges—such as quantization, VRAM/RAM allocation, and inference speed optimization—whereas using a cloud API requires nothing more than entering an API key.

There are only two scenarios where the local model approach is truly suitable: when data must absolutely remain on the local machine (e.g., processing sensitive business data) or when fully offline operation is required (e.g., within an internal network). Unless one of these conditions applies, the framework-based approach is the more rational choice.

Configuration Selection Based on Usage Intensity

For users opting for the framework mode, the 2 vCPU/2GB RAM configuration is the "sweet spot"—it offers a reasonable price and sufficient performance, ensuring the Agent isn't killed by the system due to insufficient memory while executing complex tasks.

Why is a VPS better than a local computer for running the Hermes Agent?

The biggest issue with running the Hermes Agent on your own computer is that the Agent stops whenever the computer is turned off. One of Hermes' core values ​​is "persistent cross-session memory"—it needs to run continuously in the background to accumulate skills, organize memories, and execute scheduled tasks. If it only runs for a few hours a day while the computer is on, these capabilities are significantly compromised.

A VPS provides a dedicated runtime environment that is online 24/7. Whether you are sleeping, commuting, or in a meeting, the Agent keeps working in the background. Furthermore, VPS network environments are generally more stable than home broadband, offering lower latency when accessing cloud APIs.

Jtti Hong Kong Node: A Stable Foundation for the Hermes Agent

Running the Hermes Agent in framework mode places specific demands on the VPS: stable connectivity, low latency, and zero packet loss. Since the Agent requires continuous calls to cloud APIs, frequent network timeouts or packet loss would severely degrade the conversational experience. Jtti’s Hong Kong cloud servers feature direct connectivity across all three major Chinese carriers via CN2 GIA lines: China Telecom uses CN2, China Unicom uses AS4837, and China Mobile uses CMI. Real-world tests show latency of 30–50ms for users in mainland China. This level of latency is ideal for calling APIs such as Kimi and Claude, ensuring that AI agents respond with virtually imperceptible network lag.

In terms of specifications, Jtti’s standard 2-core/4GB plan (featuring 5Mbps CN2 bandwidth and unlimited data for $29.36/month) is a sensible choice for running Hermes Agent. The 2-core/4GB configuration offers more headroom than the 2-core/2GB option, preventing memory bottlenecks during concurrent tool calls or extended, long-context conversations. For tighter budgets, the entry-level 1-core/1GB plan ($38/year) can handle basic functions, though users might experience lag in multi-session scenarios.

For applications targeting users in mainland China, the network advantages of Jtti’s Hong Kong nodes are unmatched by standard international connections. Furthermore, the policy of maintaining the same price upon renewal means that once you secure a configuration suitable for running agents, long-term operating costs remain stable and won't spiral due to market price hikes.