For anyone following Hermes Agent, this week tastes different from the usual. It is not yet another feature bolted onto the core, but the moment the project starts behaving like infrastructure: frontier models arriving in the catalogue, a release pipeline that settles down, and a set of tools for governing the complexity generated by its own growth.

A frontier model, free to try

Nous Research has made StepFun’s Step 5 Preview available on Nous Portal, free for a week. The numbers explain why the news matters: a sparse mixture-of-experts architecture with roughly 600 billion total parameters and 27 billion active per token, a context window of one million tokens and visual input as well. On the Hermes Index — the ranking Nous builds by running every model through the same agent harness — the initial score of 33.89 was later corrected to 32.83, just behind GPT-6 Luna (33.89) and GLM 5.3 Flash (34.95). List prices remain once the promotion ends: 1 dollar per million tokens on input and 2.70 on output. The move says something broader: the agent is no longer just the shell, but the test bench on which labs measure their own models.

The release learns to be stable

On 8 October came v0.21.6, the first cut by the new stable release pipeline: a single build reference, a tested Docker image and a receipt tag at publication. It packs about 2,100 pull requests accumulated since the previous version and three groups of security fixes, among them the hardening of dashboard authentication. The full curated notes will arrive with v0.22.0. It is the signal that the project wants to stop being a permanent building site.

Governing many agents, not one

On the operational side, attention shifts from the single conversation to coordination. Hermes Kanban is a persistent board, shared across profiles, where every task is a card with an assignee and a status; a dispatcher claims ready cards and starts workers in separate workspaces, with heartbeats and recovery of orphaned tasks. Around the official project a third-party ecosystem is also growing: Ares, a workbench for those who want explicit control over releases, and HOL Guard, a local security layer that inspects commands and plugins before execution. They are independent initiatives with very little traction — worth remembering — but they point to where demand is moving: security and supervision, no longer just capability.

Conclusion

The open question is no longer whether Hermes can do difficult things, but whether it can stay comprehensible while doing them. A 600-billion-parameter model free to try, a release that settles down and a board coordinating dozens of workers tell the same transition: from a powerful tool to a platform others build on. The hard part is starting now.

Verified sources: nousresearch.com e X @NousResearch (8-9 ottobre 2026); alphasignal.ai e aipromonow.com su Step 5 Preview e Hermes Index (8-9 ottobre 2026); newreleases.io/project/github/NousResearch/hermes-agent/release/v0.21.6, github.com/NousResearch/hermes-agent/releases e hermes-agent-lab.com/releases (8 ottobre 2026); hermes-agent.nousresearch.com/docs — kanban (9 ottobre 2026); X @HermesWatcher, @RecursiveIntell, @NFTerramike (9-10 ottobre 2026).