Published: 2026-07-21

Running Kimi K3 as the Model Behind Hermes: Setup, /learn Skills and Shared Memory

Chapters / key moments (click to jump — plays here on the page)

Julian Goldie swaps the model behind Hermes over to Kimi K3 (Moonshot) and walks through what changes: the provider-selection flow, saving reusable skills with /learn, and running a memory layer that both the Kimi CLI and the Hermes agent read from. The reproducible mechanics are summarised below; the video itself also demonstrates several bespoke workflows built on top of a paid setup, which we have not verified.

Source video

"Kimi K3 + Hermes Changes AI Agents Forever" by Julian Goldie SEO — Watch on YouTube →

Key Takeaways

Switching Hermes to Kimi K3 is a provider selection, not a reinstall. Run hermes model in a terminal, pick Moonshot from the dropdown, and authenticate with a coding plan. Hermes stays the agent layer; Kimi K3 becomes the reasoning model behind it.
There is a second route via the dashboard. If you manage models through a dashboard rather than the CLI, open the models list and search for Kimi. Goldie notes the entry may not appear until you hit refresh — worth knowing before assuming the provider is unavailable.
/learn converts a document into a stored skill. Passing /learn a link makes Hermes read the guide, report what it extracted, and save the result locally as a skill it can reuse in later sessions. This is the difference between a one-off answer and a capability that persists. See our Hermes skills guide for writing skills by hand.
A shared memory layer spans the CLI and the agent. Because both read from the same store, context written in the terminal is visible in the agent and vice versa, so sessions do not start cold. Our Hermes memory guide covers how that layer is configured.
MCP connects Hermes to non-coding tools. The video bridges Kimi K3 to Blender over MCP to generate 3D assets — a useful illustration that MCP servers are not limited to developer tooling. See Hermes MCP tools.
Mixture-of-agents is offered as a cost tactic. Running two open models (Kimi K3 and GLM 5.2) that critique each other is pitched as a way to avoid burning quota on closed frontier models. Treat the quality claim as anecdotal — no benchmark is shown.

Commands & Code Mentioned

hermes model          # open the provider picker, then choose Moonshot (Kimi)
/learn <url>          # read a guide and save it as a reusable local skill

What we could not verify

The video's benchmark claims — Kimi K3 ranked first on a front-end code arena and top-three on intelligence indexes — are stated on screen without a linked source, and the head-to-head quality comparisons against Fable 5 and GPT-5.6 are the presenter's own informal tests. Our benchmarks page tracks published leaderboards where these placements can be checked independently. Much of the video also demonstrates custom workflows distributed through the presenter's paid community; those are not reproducible from the video alone and are excluded from the takeaways above.