Wiki · Start here
Choosing an AI model
An agent is only as good as the model you give it. Here is how to choose without the theory — what a model must support, what we recommend and what it costs.
hermes model) and can change it any time later.What a model must be able to do
Not every model suits an agent. Whatever you pick, it must meet three conditions:
| Requirement | Why |
|---|---|
| At least 64k tokens of context | The agent packs tools, guides and memory into the conversation. Smaller context is rejected by Hermes — it would not fit the tools |
| Tool calling support | Without it the agent cannot run commands or read files — it would only talk |
| Sensible price per million tokens | The agent sends a lot of text (tools, results, history). With expensive models this adds up fast |
Reading images (vision) is useful but optional — without it the agent handles everything except photos and screenshots.
Three routes to a model
| Route | How it works | Best for |
|---|---|---|
| Nous Portal | One subscription and 300+ models under a single account. In Hermes you sign in with hermes setup --portal | The easiest start — one login, no API keys |
| API key at a provider | You create an account (OpenAI, Anthropic, Google, DeepSeek…) and pay for usage, usually by topping up credit | A specific model or the lowest price |
| OpenRouter | One account and one payment for 300+ models from different companies — switch between them with no new signups | Trying several models and comparing |
| Local model (Ollama) | The model runs on your own machine and sends nothing out | Only on a powerful computer with a GPU — unrealistic on a small VPS |
Where to get the models
Click a tile to go straight to the signup page:
The logos are monochrome provider marks (the simple-icons set) — they are here for recognition only and imply no partnership or endorsement.
What we recommend (autumn 2026)
| For | Recommendation | Why |
|---|---|---|
| Getting started and everyday tasks | DeepSeek (deepseek-flash) | Among the lowest prices, 1M token context, supports tools and images. Ideal for learning |
| Best value for money | GLM (Z.ai), Kimi (Moonshot), Qwen | Marked as recommended or default in Hermes — they handle tools well |
| Complex tasks and coding | Claude (Anthropic) or GPT (OpenAI) | Most reliable on long procedures, but several times more expensive |
| Trying it for free | models tagged :free on OpenRouter | Cost nothing, but are slower and rate limited — fine for playing, not for production |
| Sensitive data must stay in-house | a local model (Ollama) on your own machine | Data never leaves you; you do need a proper GPU machine |
What it costs
Prices are charged per million tokens (a token is a piece of a word). The figures below come from the official price lists as of 15 Oct 2026 — they change, so check before you pay:
| Model | Input (per million tokens) | Output | Note |
|---|---|---|---|
deepseek-flash | $0.15 off-peak / $0.30 peak | $0.60 / $1.20 | 1M context, peak hours cost double |
deepseek-v4-pro | $0.66 / $1.32 | $1.98 / $3.96 | stronger variant of the same family |
| models on OpenRouter | from ~$0.14 | from ~$0.28 | prices vary widely across 300+ models |
| Claude, GPT (flagships) | single-digit dollars | single to double-digit dollars | an order of magnitude above DeepSeek |
What "peak" means at DeepSeek: weekdays 01:00–04:00 and 06:00–10:00 UTC. In Czechia that is 3:00–6:00 and 8:00–12:00 (an hour earlier in winter). Off-peak and at weekends you pay half — scheduling the agent to work at night saves money.
What about a monthly bill? As a rough guide: ordinary personal use (dozens of questions a day) stays within tens of Czech crowns a month on DeepSeek. With flagship models (Claude, GPT) expect ten times that or more. The more the agent works on its own (scheduled tasks, data collection), the more it spends.
Two common misconceptions
| Misconception | Reality |
|---|---|
| "I have a ChatGPT Plus / Claude Pro subscription, so this is free." | No. Those subscriptions cover their chat in a browser. An agent needs API access, billed per usage — a different account and a different payment. The exception is Nous Portal, where the subscription includes both the model and the tools |
| "The cheapest model is enough, it does not matter." | A model with a small context will not even start, and a model without tool support will run nothing for you. The money you saved returns as hours of lost work |
How to set the model
Inside a running chat you switch models with the /model slash command — it applies immediately. Settings made from the command line apply to new conversations.
Tips from running it
| Tip | Why |
|---|---|
| Start with a cheap model (DeepSeek) | You learn what the agent can do for a few crowns. An expensive model only pays off on tasks the cheap one cannot handle |
| Reach for a flagship only on hard tasks | Switch with /model mid-task — you then pay the expensive model only for what it really has to do |
| Set an auxiliary model too | Side tasks (reading images, summaries) can go to a cheap model while the main one stays high quality |
| Schedule recurring jobs off-peak | DeepSeek is half price off-peak, and a night job bothers nobody |
| Watch how full the context gets | The /context command in a chat shows how full it is; /compress trims it |