POPA ReconHermes: The Moat Above The Model
Hermes: The Moat Above The Model Open and closed models have converged on benchmarks. Inference cost collapsed from $30/1M tokens to ~$0.05. So the model no longer differentiates. What matters is the layer above it: memory, skills, context management, error recovery. Nous open-sourced a runtime like Claude Code. Memory and skills stay local and user-owned. It went from 4th to 1st on OpenRouter in eight weeks. It now routes ~3.2T tokens/week. The claimed loop: sessions generate behavioral data, that data improves post-training, better models attract more users. Revenue comes from tiered Nous…
The excerpt is quoted from a public Telegram channel, which owns its text.