Using DeepSeek with Hermes Agent: setup, models and prices

Two ways to connect DeepSeek

• A DeepSeek API key. Hermes has a built-in deepseek provider that talks to DeepSeek directly. You create a key on DeepSeek’s platform and pay DeepSeek per token, separate from any chat app. This is the route DeepSeek’s own Hermes guide describes.
• OpenRouter. One OpenRouter key reaches DeepSeek and many other models, so you can try DeepSeek next to the others without a new account. Hermes’ model list for OpenRouter includes deepseek/deepseek-v4-pro and deepseek/deepseek-v4.1-flash.

Hermes also lists DeepSeek models on some other gateways, but these two are the plain routes. Running DeepSeek models on your own hardware is a different setup and is not covered here.

Setting it up on your own machine

If you have not installed Hermes yet, start with how to install it. Then run hermes model in a terminal, choose DeepSeek, paste your key and pick a model. Alternatively, run hermes setup model, choose DeepSeek, enter your key and pick a model. The choice is saved as model.provider and model.default in ~/.hermes/config.yaml. After you pick a model, the wizard also asks for a reasoning effort.

To do it by hand instead, save the key (Hermes puts API keys in ~/.hermes/.env) and name the provider:

hermes config set DEEPSEEK_API_KEY sk-...
hermes chat --provider deepseek --model deepseek-flash

To make it the default, set it in ~/.hermes/config.yaml:

model:
  provider: deepseek
  default: deepseek-flash

The key variable is DEEPSEEK_API_KEY. There is an optional DEEPSEEK_BASE_URL for a different endpoint, which you will normally leave alone. Once DeepSeek credentials are in place, by hermes model, hermes setup or the manual steps above, /model switches between that provider and others inside a conversation.

You can also add a safety net: hermes fallback adds a second provider that Hermes switches to when the first one fails or is rate-limited, written to fallback_providers in the same config file. DeepSeek as the main model with a different provider as the fallback is a reasonable pairing.

Which DeepSeek model to pick

Hermes’ model picker for the DeepSeek provider shows two names, and DeepSeek’s documentation lists the same two on 23 October 2026: deepseek-flash and deepseek-v4-pro. DeepSeek’s announcement of V4.1 Flash (10 September 2026) says that from 14 September all deepseek-v4-pro requests are routed to V4.1 Flash at Flash rates, and that this continues until a V4.1 Pro launches. The pricing page still lists a separate Pro price, so the two official pages disagree. Until that is resolved, pick deepseek-flash and check DeepSeek’s pricing page before you rely on Pro being a different model.

• deepseek-flash (currently DeepSeek-V4.1-Flash). A 1 million token context, up to 384K tokens of output, tool calls, thinking mode and vision. DeepSeek reports it ahead of V4-Pro on tests by several parties. A good fit for scheduled jobs, inbox sorting, research, summaries and anything that runs many times a day. Hermes also uses it as the model for small side tasks when DeepSeek is your provider.

One thing to fix if you have used DeepSeek before: the old names deepseek-chat and deepseek-reasoner were retired on 24 July 2026, according to the Hermes source. When you use Hermes’ native deepseek provider, Hermes maps these retired IDs to deepseek-flash, but it is tidier to change them. That mapping is specific to the native provider; through OpenRouter or a custom endpoint, use the model name that route lists. DeepSeek says the older deepseek-v4-flash name is still accepted and is served by the current Flash model.

What it costs

DeepSeek’s price for deepseek-flash per million tokens on 23 October 2026. Off-peak is half the peak price. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays; every other hour is off-peak.

This page leaves out the separate deepseek-v4-pro prices that DeepSeek’s pricing page also shows, because its announcement says Pro requests are billed at Flash rates for now. If you choose Pro, check which price DeepSeek applies.

A rough feel for one API request: 20,000 input tokens with no cache hits and 1,000 output tokens come to $0.0072 at peak prices and half that off-peak. Output tokens include any thinking the model does. Cache hits can reduce input costs, and DeepSeek’s caching is best-effort, not guaranteed. A complete agent turn may make several billable requests, so multiply accordingly.

DeepSeek also sells a separate search and fetch API, at $0.75 per 1,000 successful requests. That is not part of the model price above, and Hermes does not need it to work.

Thinking mode and tool calls

Both DeepSeek models can think before they answer. DeepSeek turns thinking on by default, with a default effort of high. Hermes does not leave this to chance: for DeepSeek it always sends the thinking setting explicitly, so the on or off choice and the effort level you set in Hermes are respected.

There is a technical reason behind that. When a thinking-mode request includes tools, DeepSeek requires the reasoning from earlier turns to be passed back to it, and it returns an error if it is missing. Hermes keeps and replays that reasoning for DeepSeek automatically. It recognises DeepSeek by the provider name, by the endpoint address (api.deepseek.com) or by a model name containing deepseek. Only if none of those identifies an endpoint that needs the replay, such as a gateway that renames the model, set reasoning_echo: true on that model: entry.

• In the terminal, /reasoning high changes the effort for the session, /reasoning high --global keeps it, and /reasoning none turns thinking off. You can also set agent.reasoning_effort in the config file. Hermes passes low, medium, high and max to DeepSeek. DeepSeek’s documentation says it folds the requested levels into fewer real ones, for example treating medium as high.
• In thinking mode DeepSeek ignores temperature, presence_penalty and frequency_penalty, so changing them has no effect.
• DeepSeek’s native API supports JSON mode but not strict JSON-schema output, and Hermes accounts for that.
• If a model sometimes describes an action instead of doing it, agent.tool_use_enforcement adds a stricter instruction. It is already on automatically for model names containing deepseek.

For a cheap always-on agent, a sensible setup is Flash with the effort left at its default, raising the effort for a job when you see it getting something wrong.

DeepSeek on ReadyHermes

ReadyHermes runs your own Hermes Agent on our servers, always on, so the schedule and your chat apps keep working while your computer is off. DeepSeek shows up in two places:

• The included AI. Every paid plan includes US$1 of AI each month on DeepSeek V4.1 Flash, roughly 100 everyday requests, renewing on your billing date. The free trial hour runs on its own included AI route.
• Your own DeepSeek key. DeepSeek is one of the providers in the workspace’s Model key tab, next to OpenRouter, OpenAI, Anthropic, Google Gemini and others, on paid plans. Choose DeepSeek, paste your key, press Load models to see what the key can use, pick a model such as deepseek-flash, then press Check and switch. We send one tiny test message first so a typo cannot leave your agent silent, and the agent then restarts on your key.

Your own key is billed by DeepSeek directly, and ReadyHermes adds no AI limit on it. Use included AI switches back. ReadyHermes does not run models on your own computer or host GPUs for models you run yourself. If you want to compare running Hermes yourself with a hosted agent, see self-hosting vs managed hosting, or start with what Hermes Agent is.