Guide
Hermes Agent + Mistral: Use Mistral AI's Open and Hosted Models
Mistral — the Paris-based AI lab's model family — pairs frontier quality with a strong open-source story. Mistral Large 2 handles deep reasoning, Mistral Medium 3 is the balanced workhorse, and Codestral leads on coding benchmarks for its size class.
What Is Mistral?
Mistral AI is a European AI lab known for shipping strong open-weight models alongside hosted commercial offerings. Mistral pioneered mixture-of-experts architectures with Mixtral and has consistently delivered top-tier multilingual performance — particularly strong on French, German, Spanish, and Italian.
Hermes Agent reaches Mistral through three paths: La Plateforme (Mistral's hosted API), OpenRouter for unified access, or self-hosted for the open-weight variants (Mistral 7B, Mixtral 8x7B/8x22B, Codestral Mamba).
Mistral Model Lineup for Hermes
| Model | Best For | Context | Cost (Input) |
|---|---|---|---|
| Mistral Large 2 | Frontier reasoning, complex tool use | 128K tokens | ~$2/M tokens |
| Mistral Medium 3 | Everyday agent default — balanced | 128K tokens | ~$0.40/M tokens |
| Mistral Small 3 | Fast chat, high-volume bots | 128K tokens | ~$0.20/M tokens |
| Codestral | Code generation, refactoring, FIM | 32K tokens | ~$0.30/M tokens |
| Mixtral 8x22B (open) | Self-hosted MoE, strong multilingual | 64K tokens | Self-host cost only |
Option 1: Hermes Agent on OpenClaw Launch (Easiest)
- Go to openclawlaunch.com/hermes-hosting and start a Hermes deploy.
- Select Mistral Medium 3 (or Large 2 / Codestral) from the model dropdown.
- Connect Telegram, Discord, WhatsApp, or another channel.
- Click Deploy. Your Mistral-powered Hermes Agent is live in roughly 30 seconds.
Option 2: Mistral La Plateforme API Direct (Self-Hosted)
# Hermes reads MISTRAL_API_KEY
export MISTRAL_API_KEY=...
hermes inference set mistral
hermes model set mistral-medium-3
# config.yaml equivalent:
# inference:
# provider: mistral
# model:
# default: mistral-medium-3Generate an API key at console.mistral.ai/api-keys. Billing is usage-based with EU data residency available.
Option 3: Mistral via OpenRouter (Self-Hosted)
export OPENROUTER_API_KEY=sk-or-...
hermes inference set openrouter
hermes model set mistralai/mistral-large-2Option 4: Self-Hosted Mistral via Ollama or vLLM
Mistral 7B, Mixtral 8x7B, Mixtral 8x22B, and Codestral are open-weight under Apache 2.0. Run them locally for full data privacy.
# Ollama path (easiest)
ollama pull mixtral:8x22b
export OLLAMA_HOST=http://127.0.0.1:11434
hermes inference set ollama
hermes model set mixtral:8x22b
# vLLM path (production)
vllm serve mistralai/Mixtral-8x22B-Instruct-v0.1 --port 8000
export OPENAI_API_BASE=http://127.0.0.1:8000/v1
hermes inference set openai
hermes model set mistralai/Mixtral-8x22B-Instruct-v0.1When to Choose Each Mistral Model
Choose Mistral Medium 3 as your default. It hits roughly Claude Sonnet quality on tool use and reasoning at a fifth of the cost — the sweet spot for everyday Hermes deployments.
Choose Mistral Large 2 for complex multi-step agent tasks where you'd otherwise use Claude Opus or GPT-5.5. Strongest on European languages.
Choose Codestral for coding agents: strong fill-in-the-middle, 80+ programming languages, and dedicated FIM tokens that pair well with skill-based workflows.
Choose open-weight Mixtral when you need EU data residency or fully air-gapped deployment — the Apache-2.0 license permits any use.
Switching Models at Runtime
/model mistralai/mistral-medium-3
/model mistralai/mistral-large-2
/model mistralai/codestralWhat's Next?
- Hermes Agent + Llama — Another open-weight option
- Hermes Agent + Claude — Anthropic Claude with Hermes
- Hermes Agent + vLLM — Production self-host for Mistral
- Hermes Agent + OpenRouter — One key for 200+ models