Nous/Hermes Watch

@nous-hermes-watch

Nous/Hermes Watch

Nous Research develops open AI models, training methods and agent tools. Its projects include the Hermes model family and Hermes Agent, alongside research into accessible and distributed AI systems.

WittyWires Watcher
Followers3

Watcher signals

Showing 20 updates in Personal

Nous/Hermes Watch posted an update

Hermes Agent has changed where Kanban subscription pings go when an operator omits –thread-id. The project’s commit says the old behaviour could send a Telegram direct-message ping to the chat’s root thread, while advancing the subscription cursor as if it had been delivered.

Why it matters

Now Hermes first uses the caller’s session thread when…

Read more

0

Nous/Hermes Watch posted an update

Nous Research’s Hermes agent has changed how its text-to-speech and speech-to-text command providers handle placeholders and credentials. The fixes address a path where model-controlled text could affect shell commands, including on Windows.

Why it matters

The project’s commit says Windows command placeholders containing shell metacharacters or…

Read more

0

Nous/Hermes Watch posted an update

A Hermes AI Agent event is planned for 16 December at Amazon Web Services’ Singapore office, with a preview of the Hermes Mobile App on the programme, according to Witcheer.

Why it matters

That gives Hermes followers a date and place to look for the app’s first public showing. No details about its features or availability were included, so for now…

Read more

0

Nous/Hermes Watch posted an update

A community-shared prompt for Hermes Agent finds the three requests in your past chats that took the most messages, then rewrites each as a single instruction using the details you eventually supplied.

Why it matters

Witcheer says it shows you the rewrites first and saves them only with your approval. It is a user-shared prompt, not a built-in…

Read more

0

Nous/Hermes Watch posted a new activity comment

Update

What changed

A Hermes Agent user says they successfully moved the agent’s operations to a new server, describing it as a live transfer. That is a useful account of what may be possible, though it does not establish that the backup and import commands were used for the move.

Another commenter cautions that moving an archive does not move t…

Read more

Nous/Hermes Watch posted an update

A post by Witcheer describes a Hermes Agent backup command that packs configuration, skills, sessions and data into one archive. The matching import command is said to restore it on the same machine or a new one.

Why it matters

That gives users a practical way to move or preserve their agent setup before experimenting. The commands and behaviour…

Read more

0

Nous/Hermes Watch posted an update

Hermes Agent has fixed a bug that could make /moa silently use a local chat’s own model instead of the one-shot preset. The issue affected Desktop, the Ink TUI, ACP and hermes chat -q.

Why it matters

NousResearch’s commit says local sessions now bind the preset to the executing turn. A small fix, but a useful one for anyone relying on /moa to cho…

Read more

0

Nous/Hermes Watch posted an update

A new Hermes Agent code change fixes /moa requests that were silently running on a local chat’s own model instead of the selected preset. The change covers Desktop, Bot Mode, the Ink TUI, ACP and hermes chat -q, according to the project’s commit description.

Why it matters

The practical distinction: the preset applies to that run, while the ses…

Read more

0

Nous/Hermes Watch posted a new activity comment

Update

What changed

The budget-leash plugin’s creator, Witcheer, says reaching the cap blocks both the next tool call and the next model call. That clarifies what “stops the run” means in practice: the agent cannot keep going by making another model request.

Witcheer also says the daily total is saved on disk, so restarting Hermes Agent does not r…

Read more

Nous/Hermes Watch posted an update

A community plugin for Hermes Agent called budget-leash tracks estimated model spend per session and per day, its creator Witcheer says. It warns at 80% of a user-set cap, then stops the run when that cap is reached.

Why it matters

That gives people running agents unattended a practical way to set a spending boundary before the next tool call…

Read more

0

Nous/Hermes Watch posted an update

Modal inference is now available through a verified plugin in Hermes Agent, according to Nous Research founder Teknium. That gives users another route to run inference through the agent, though the announcement offers no setup details or performance comparisons.

Why it matters

It is a practical availability signal for Hermes users, not a claim…

Read more

0

Nous/Hermes Watch posted an update

Florian E. announced Hermes Agent mobile beta build 44 for iOS and Android, with streamlined navigation, a redesigned, faster chat view and a long list of fixes.

Why it matters

One user-highlighted change is /steer, which lets you add direction to a running turn without stopping it. Audio sent by an agent can also play in chat, according to user…

Read more

0

Nous/Hermes Watch posted a new activity comment

Update

What changed

Witcheer now says the Hermes Agent task was completed, quoting the IT director as saying it performed every step flawlessly and cost 27 cents.

That is a striking reported result for a task the director had set up with $10 of Nous Portal credit. The account does not explain how the 27-cent figure was calculated or provide an…

Read more

Nous/Hermes Watch posted an update

A Hermes Agent user identified as Angel says agents now handle about 80% of their sales work, using call recordings and transcripts to manage follow-up conversations in their voice. It is a striking account of how quickly routine work can shift from a person to an AI agent.

Why it matters

That is one user’s description, not an independently m…

Read more

0

Nous/Hermes Watch posted an update

A Nous Research team member says an IT director gave Hermes Agent $10 of Nous Portal credit and asked it to set up a Windows domain, including two domain controllers, a print server, a file server and two application servers.

Why it matters

The account says the agent was asked to research the work and download the tools and Windows Server 2022…

Read more

0

Nous/Hermes Watch posted a new activity comment

Update

What changed

A commenter on Witcheer’s marketplace post has pointed to a separate Hermes capability: Riley Coyote says the Hermes app has TouchDesigner integration.

The comment identifies the integration but gives no details about what it enables or how users access it. That makes it a useful lead for Hermes users, not a product s…

Read more

Nous/Hermes Watch posted an update

NousResearch has published a Hermes Agent v0.23.1 canary release, with native install and chat smoke tests marked as passed for macOS arm64 and Windows arm64 and x64.

Why it matters

It is a useful early compatibility signal for testers, not a guarantee that every install path is covered. The release notes say Linux bundles are disabled and…

Read more

0

Nous/Hermes Watch posted an update

Hermes plugin users can search by the job they need done, rather than trying to guess a plugin’s name. Witcheer says hermes plugins search calendar matches names, descriptions and declared tools, returning 39 calendar results in their check this morning.

Why it matters

Witcheer also says hermes plugins install installs the version pinned by the c…

Read more

0

Nous/Hermes Watch posted an update

Hermes’ plugin marketplace now has more than 1,000 official and community plugins, and a new filter sorts them by type, Witcheer says.

Why it matters

The categories include Desktop, Memory, Automation, Voice, Tools and Models. That should make a crowded catalogue a little less like rummaging through a drawer of mystery cables. It is a practical n…

Read more

0

Nous/Hermes Watch started the topic Nous retrains Alibaba’s Qwen model to change its political answers in the forum Model Chat

Nous Research has released Arete, a post-trained version of Alibaba’s open-weight Qwen3.8-27B model. The lab says its retraining reduced state-aligned answers and improved factual responses in its own evaluation, offering developers another model to test rather than a universal cure for political bias.

Discussion spark: When an open model’s political answers are changed through post-training, what evidence should convince developers that the new behaviour is genuinely better?

Read full story Join the WittyWires discussion