Hermes recipe

Hermes Agent and Hermes Desktop: how to get one real DeepSeek V4 request through first.

Most Hermes plus DeepSeek failures fall into three buckets: provider-specific request shaping, an HTTP versus HTTPS mismatch, or stale model ids. Start by separating the raw endpoint from the framework adapter.

Current DeepSeek V4 names: start with deepseek-v4-flash for the first test. Use deepseek-v4-pro after the adapter works. Legacy aliases deepseek-chat and deepseek-reasoner are scheduled to retire on July 24, 2026.

1. Run the smallest useful direct request

If this succeeds, your key and model are fine. The next bug is in the Hermes-side adapter path.

curl https://api.deepseek.com/chat/completions \
  -H "Authorization: Bearer sk-your-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [{"role": "user", "content": "hello"}]
  }'

2. If Hermes built-in DeepSeek returns HTTP 400 on the first message

This usually means the built-in provider path is adding or normalizing request fields in a way the current DeepSeek endpoint rejects. A practical workaround is switching to the custom OpenAI-compatible path first.

model:
  default: deepseek-v4-flash
  provider: custom
  base_url: https://api.deepseek.com
  api_key: sk-your-key
  api_mode: openai-completions

Do the first test with one plain user message. Leave tools, streaming, long context and reasoning-specific controls for later.

3. If Hermes Desktop shows HTTP 307 Temporary Redirect

That points to an HTTP versus HTTPS mismatch. Even if your config looks right, inspect the final request host in the log. If the client is sending http://api.deepseek.com/v1, CloudFront will redirect and the client may not follow it.

Check the final runtime URL, not only the saved config.
Make sure the resolved endpoint is https://api.deepseek.com or https://api.deepseek.com/v1.
Retest with one direct chat completion before retrying the desktop flow.

4. Debug in this order

Direct /chat/completions request works.
Hermes sends the exact model id you expect.
Hermes sends HTTPS, not HTTP.
Only after that, test tools, thinking mode, or multi-turn state.

5. Quick route choice

Situation	Best next step	Why
Direct DeepSeek request already works	Fix the Hermes adapter path	The provider key and model are not the blocker.
You only need one hosted OpenAI-compatible endpoint test	Use a narrow prepaid trial path	Useful when signup, payment, or provider setup overhead is the real delay.
You need larger model coverage later	Expand after the first request succeeds	Do not debug five variables at once.

6. If the access path itself is the blocker

If your team is stuck on account setup, payment rails, or just needs one small OpenAI-compatible DeepSeek V4 test before spending more time, Black Eagle AI keeps the first step intentionally small.

See the $5 starter flow Open integration recipes Read the billing FAQ

One small test first

Register, top up a small amount, and validate the OpenAI-compatible path before treating Chinese model access as a bigger migration project.

Create account