Using Mistral models with Hermes Agent: setup, models and prices

Does Hermes Agent have a built-in Mistral provider?

For chat, no. In version 0.22.5 the provider list behind hermes model has no Mistral entry, unlike DeepSeek, which has its own (see Hermes Agent with DeepSeek). Hermes’ own provider documentation lists Mistral AI among the services that work through its custom endpoint, with the base address https://api.mistral.ai/v1, because Mistral’s chat API speaks the same OpenAI-style protocol Hermes already understands.

Mistral does appear in Hermes in one built-in place: speech. Voice-note transcription and text-to-speech both have a Mistral option that reads MISTRAL_API_KEY, covered in the Voxtral section below.

So you have two practical routes for the chat model:

• A Mistral API key as a custom endpoint. You pay Mistral directly and can use any model your key can reach.
• OpenRouter. One OpenRouter key reaches Mistral models next to many others, so you can try them without opening a Mistral account.

A third route, running open-weight Mistral models yourself, is covered further down.

Connecting the Mistral API as a custom endpoint

If you have not installed Hermes yet, start with how to install Hermes Agent. Create an API key in Mistral’s console, then run hermes model in a terminal, choose the custom endpoint option, and enter the base URL, the key and the model name.

To write it by hand instead, Hermes supports named custom providers. Put the key in ~/.hermes/.env and describe the endpoint in ~/.hermes/config.yaml:

hermes config set MISTRAL_API_KEY your-key-here

providers:
  mistral:
    api: https://api.mistral.ai/v1
    key_env: MISTRAL_API_KEY

model:
  provider: custom:mistral
  default: mistral-small-2603

Inside a chat, /model custom:mistral:mistral-small-2603 switches to a model on that endpoint once it is set up. Hermes is careful about one Mistral quirk: the extra “think” field it sends to Ollama is not sent to other custom hosts, because strict servers such as Mistral reject it.

You can also add a safety net: hermes fallback add picks a second provider that Hermes switches to when the first one fails or is rate-limited. A Mistral main model with a different provider as the fallback is a reasonable pairing.

Using the same MISTRAL_API_KEY name for chat and for voice is fine: the key is the same Mistral account in both places.

OpenRouter as the other route

OpenRouter has its own provider in Hermes, so it needs no custom settings: run hermes model, choose OpenRouter and paste the key. OpenRouter’s Mistral page lists dozens of Mistral models, including Codestral, Devstral, Magistral, Ministral, Mixtral, Pixtral and Voxtral families, and OpenRouter’s own site shows which are currently available (page read 31 October 2026).

The model list Hermes shows for OpenRouter is a short hand-picked set, and I found no Mistral model in it. That does not block you. The picker offers an “Enter custom model name” choice, so type the exact slug from OpenRouter’s page, which starts with mistralai/.

The trade-off is the usual one for a middleman: one account and one bill for many models, in exchange for OpenRouter’s own pricing and routing rather than Mistral’s. If Mistral is the only provider you plan to use, the direct key above is the simpler path.

Which Mistral models suit agent work

An agent calls tools many times per task, so what matters is reliable function calling and a large context window, then price. From Mistral’s model pages on 31 October 2026:

• Mistral Small 4 (mistral-small-2603): a general model that combines instruct, reasoning and coding, with a 256k context window, function calling and an Apache 2.0 licence. It is the sensible default for an everyday agent.
• Mistral Large 4 (mistral-large-4): Mistral’s largest open-weight general model, with a 1M context window and function calling. Its page marks it as a public preview released on 6 October 2026, so treat it as newer and less settled than Small 4.
• Ministral 3 in 3B, 8B and 14B sizes (for example ministral-8b-2512): small Apache 2.0 models with a 256k context window and function calling. They are the cheap choice and the ones to try locally.
• Codestral (codestral-2508): a code model with a 128k context window, built for fill-in-the-middle and code generation. It fits editor completion better than a general agent that also writes messages and manages a schedule.

Mistral’s model overview lists Large 4 and Small 4 under models for broad reasoning, coding, tool use and agentic tasks. As with any model, run your own real tasks before you settle. Hermes normally refuses to start with a model whose context window is under 64,000 tokens, and every model above clears that easily; the floor matters mostly for local setups, below.

What it costs, and the free tier

Mistral charges per million tokens, input and output counted separately. The prices on each model’s page on 31 October 2026 were:

• Mistral Small 4: $0.15 per million input tokens and $0.60 per million output tokens.
• Mistral Large 4: $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.
• Ministral 3 8B: $0.15 per million tokens; Ministral 3 14B: $0.20 per million tokens.
• Codestral: $0.30 per million input tokens and $0.90 per million output tokens.

Mistral’s pricing page adds that batch processing cuts the price by 50% and cached input tokens by up to 90%. It also quotes a different example figure for “Mistral Large” in its FAQ than the Large 4 model page does, so read the page of the exact model you choose. A complete agent turn can make several requests, because each tool call is another round trip, so estimate from real usage rather than from one message.

Free tier. Mistral’s usage-and-limits documentation says free mode lets you create API keys and use an included monthly allowance within the limits shown on the Limits page of the admin panel, and that pay-as-you-go extends this using prepaid credits. I could not find the allowance numbers on a public page, so look at that Limits page in your own account. Do not confuse this with the Free plan on Mistral’s pricing page, which is limited access to its Vibe apps and also gives access to Mistral Studio, its developer console.

For a wider look at what Hermes itself costs, see is Hermes Agent free.

Running Mistral models locally, and the 64k floor

Mistral publishes several models as open weights. Ollama’s library page for ministral-3 (read 31 October 2026) lists 3B, 8B and 14B tags of 3.0 GB, 6.0 GB and 9.1 GB, a 256K context window and native function calling. That makes it the practical Mistral family for a home machine. The full Ollama setup is in Hermes Agent with Ollama; in short, you use Hermes’ custom endpoint with http://localhost:11434/v1.

The catch is context. Hermes Agent normally requires at least 64,000 tokens of context and refuses smaller windows at startup, because the system prompt, the tool descriptions and the working conversation need room. A local server does not serve the model’s full window by default: Hermes’ documentation gives Ollama’s defaults as 4,096 tokens with under 24 GB of video memory and 32,768 with 24 to 48 GB. Raise it before you start Hermes:

OLLAMA_CONTEXT_LENGTH=64000 ollama serve

Check the result with ollama ps. If the server really serves a window of 64K or more but Hermes reads a smaller one, set model.ollama_num_ctx in config.yaml to the real value. This setting is honoured on any local endpoint, not only Ollama.

Two cautions from the same pages. Ollama’s mistral-small page shows 32K context on its newest listed tags, which is below what Hermes normally needs, so pick a model with at least 64,000 tokens of supported context. It does not list Mistral Small 4. And Mistral describes Small 4 as 119B parameters in total, so check the memory a download needs before you pull its weights. LM Studio is the other local route Hermes documents; it shows models trained for tool calling, Mistral among them, with a tool badge, and you load the model with at least 64,000 context.

Voxtral: Mistral for voice notes

Hermes can turn voice messages from chat apps into text before the agent reads them. Unless you pick a provider, it prefers a local transcription model if one is available and otherwise tries any cloud provider you have a key for. To choose Mistral explicitly, set the provider and the key:

hermes config set MISTRAL_API_KEY your-key-here

stt:
  provider: mistral
  mistral:
    model: voxtral-mini-latest

Hermes tries to install Mistral’s Python package the first time you use it, if lazy installs are allowed. A recording can be up to 500 MB or 60 minutes per request. Hermes gives the agent the transcript text only, without speaker labels. The same Mistral key also powers a text-to-speech option, tts.provider: mistral, whose audio comes out in a format chat apps play as a voice bubble without extra software.

On Mistral’s side, Voxtral Mini Transcribe 3 was priced at $0.0035 per minute on its model page, and its page marks it as a public preview released on 20 October 2026. Hermes’ sample config also names voxtral-mini-2602, which Mistral lists as deprecated from 31 October 2026, so keep the voxtral-mini-latest default and check Mistral’s page for what it points to. Voice notes depend on the chat app too; see Hermes Agent on WhatsApp and on Telegram.

Mistral on ReadyHermes

ReadyHermes runs your own Hermes Agent on our servers, always on, so your schedule and chat apps keep working while your computer is off. Mistral is not one of the named providers in the workspace’s Model key tab. That list holds OpenRouter, OpenAI, Anthropic, Google Gemini, DeepSeek, Groq, Qwen, Z.AI, Moonshot and MiniMax. Two routes get close:

• OpenRouter. Choose OpenRouter, paste your key, press Load models or type the model name, then press Check and switch.
• Other endpoint. The tab also has an Other endpoint choice described as any OpenAI-compatible API. It accepts an https address on the standard port with a public hostname, and Mistral’s https://api.mistral.ai/v1 has that shape. Enter it with your Mistral key, load or type a model name and press Check and switch. I have not run a Mistral key through this panel myself, so rely on its test message: it sends one small request first, and if Mistral refuses the key or model you see the reason and nothing changes.

Switching restarts the assistant, which takes about a minute. The Model key tab is open during the free trial and on paid plans, and a key added during the trial stays when you choose a plan. Your provider bills you directly. I did not find a Mistral speech-to-text or text-to-speech setting in the ReadyHermes workspace, so the Voxtral setup above is for Hermes you run yourself.

To compare running it yourself with a hosted agent, read self-hosting vs managed hosting.