Published: 2026-07-30

Kimi K3 Open Weights Paired With Hermes Agent

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

Julian Goldie covers Moonshot's "Kimi K3 open day" on 27 July — weights, technical report and three internal training tools released at once — and argues the case for pointing Hermes Agent at whoever hosts it rather than running it yourself, since he puts the download at roughly 1.4 TB. Most of the runtime is a pitch for his paid community's prebuilt agent bundle, so the reusable parts are the release facts and a short list of rough edges: K3 always reasons before answering (no quick mode), one report flagged a high hallucination rate in its own testing, and Hermes quietly degrades if you never configure its small side model for image reading, page summarising and chat compaction. He also runs through Hermes 0.19 "Quicksilver" (20 July): time-to-first-word down roughly 80%, approvals now checked by a separate reviewer model with user-defined deny rules, a delivery ledger that re-sends finished work after a crash, live sub-agent logs, and a migration command for former OpenClaw setups.

Figures and dates above are as stated in the video and are not independently verified here. For our own tracked release notes see the changelog.

Watch on YouTube → · ← Back to News

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Configuration shown

Kimi K3 facts as the speaker states them (unverified)

Moonshot released the full open weights on July 27, 2026 ("K3 open day"), with the technical report and three internal training tools. 2.8 trillion parameters, described as the largest open model released. 1 million token context. Handles images and video in one model. Reported first place on the front-end code arena (a blind human preference test). Download is around 1.4 TB, so it is not runnable on a laptop, gaming PC or most company servers; Together AI had it hosted on day one. Alibaba announced a 2.4 trillion parameter open model of its own the same weekend.

Hermes 0.19 "Quicksilver" (July 20, 2026), per the speaker

Time to first word down about 80%, from about 4.5 seconds to under one. Approvals: a separate reviewer model checks risky commands instead of asking you every time, and you can write deny rules that block things regardless. Delivery ledger: if Hermes finishes a job and crashes before sending the answer, it saves the answer and delivers it on restart. Every delegated sub-agent task writes its own log you can follow live.

Template shown in the video. Adapt to your project — do not copy verbatim without reviewing each line.

Reproducible steps

Teach Hermes from a page
Hermes has a "learn" command: paste a link, it reads the page and saves what it learned as a skill in a folder on your machine, so it answers the same way each time and persists after you close the window.
Point the agent at a hosted model, not a download
Because the weights are too large to self-host for most people, use a hosted provider and give the agent its name and key. The speaker's point is that the agent layer should not care which model sits behind it; switch the model in one command when something better ships.
Use K3 for heavy jobs and a quicker model for light ones
He says K3 always runs a full reasoning pass (no quick mode); independent testers logged over 13,000 thinking tokens on one short task.
Migrating
He says Hermes has a migration command that pulls over 30 categories of an old OpenClaw setup across; the command name is not given in the transcript.

Gotchas

Setup trap he warns about: Hermes uses a small side model for reading images, summarising pages and compressing long chats. If you never configure it, those features silently stop working with no error. Point the side model at your main provider in the config.
One report he cites flagged a high hallucination rate for K3 in its own testing, so check the output.
Promotional avatar video: the workflows shown (Oracle, Astros competitor tracking, Blender MCP 3D models) belong to his paid Agent OS bundle. Verify every Hermes command and the version claims against the official Hermes docs and release notes before relying on them.