Gladys Assistant 4.79: Smarter AI, Enhanced Voice Assistant and More Matter Devices

[pierre-gilles]

Hello everyone,

The v4.79 version of Gladys is out

Artificial Intelligence

This version brings several improvements to the AI agent:

• Scene creation: Schema correction + timeout increase to fix a scene creation bug reported by @GBoulvin and @Jluc.
• Better error handling: clearer feedback when the AI does not respond.
• Tool calls via Telegram: On Telegram, you can now see tool calls as on the web

• New tools:
  • Web Request: the agent can query APIs or web pages. This is a personal need that I implemented, and it’s ultra convenient!
  • Compare Times: time comparison

These new tools allow you to do very powerful things, for example, fetch the opening hours of a business and tell you if the business is currently open! Super convenient in scenes via the « Ask the AI » action.

The possibilities are endless!

---

Voice Assistant

• Stop button to interrupt an ongoing response.
• Microphone detection: If no sound has been recorded, Gladys displays an error message.

---

Matter

Extended support for new types of devices:

• Vacuum cleaners
• NO₂ sensors (nitrogen dioxide index)
• Fans

---

Zigbee2mqtt

• Update to Zigbee2mqtt 2.12.0
• Support for the SONOFF SNZB-01M remote control (4 buttons)

---

Node-RED

• Version selector to switch to a major version of Node-RED directly from the Gladys interface.

Convenient for switching to v5 which has just been released!

Be careful, make sure the modules you use are compatible with v5!

---

Scenes

• Expanded value field for sensor thresholds
• Back button to navigate more easily in the scene editor.

---

Gladys Plus

• When a remote user is deleted on Gladys Plus, the corresponding local user is automatically deleted on your instance.

Full Changelog

Release v4.79.0 · GladysAssistant/Gladys

As always, the update is automatic within 24 hours if you use Watchtower, or you can do it manually by going to « Settings » → « System ».

Happy updating everyone!

[mutmut]

Little test via Telegram and there seems to be an issue, it seems

And the AI creating a scene without me asking, I find that more than questionable

EDIT: in the end, no scene was created

[pierre-gilles]

Thanks for the feedback! I’ll try to reproduce it. Sometimes the AI doesn’t feel the need to call the tools at its disposal. I think the fact that it’s a « pump » isn’t a word that fits in the lexical field of things to turn on/off like a light, that must play a role

[mutmut]

The AI worked perfectly fine in 4.78:

However, tonight before upgrading to 4.79, the AI did some weird things and stopped working, even though at 1:00 PM I was able to start it without any issues and no debug logs:

[mutmut]

There are some weird things with the AI’s answers today or maybe my question wasn’t clear

[pierre-gilles]

I don’t think this is related to 4.79, I also made changes on the Gladys Plus server side!

Before I was using exclusively Gemma 4 as a model, and now I’m using a mix of Gemma 4 for image analysis, and Mistral Small 3.2 for text.

The reason is that Gemma 4 hallucinates quite a bit, more than Mistral, but Mistral is less good at analyzing images.

I think I’ll have to retest with other models to find a more versatile model

I need to create a dataset of examples to run my tests and evaluate the performance of a model in the case of Gladys.

Regarding your feedback on Matter, I don’t have the answer right now I’ll let you know when I’ve had a chance to look into it!

[pierre-gilles]

Thanks @mutmut and @Nagromdark for your feedback on the Matter integration, could one of you provide me with Gladys logs after trying to click on « update » (downloadable in one click from « Settings » → « Systems »)

@Nagromdark

I also detected another issue with my water tank level sensor, it switched to meters:

Hi @Nagromdark, this seems to come from Zigbee2mqtt, I asked the AI to investigate and:

Since Zigbee2MQTT 2.11.0, a regression was introduced in the Tuya converters for water level sensors (ME201WZ, ME202WZ, 872WZ, etc.):

• Units got mixed up between mm, cm, and m
• The installation_height and liquid_depth_max parameters changed units without the values being properly converted
• Users saw anomalous values (e.g., ~10 m instead of ~1 m)
• Several confirmed that reverting to 2.10.1 fixed the issue

Open issues:

• zigbee2mqtt#32169 — 872WZ sensor
• zigbee2mqtt#32228 — ME201WZ

A fix was merged on June 13, 2026, in zigbee-herdsman-converters (#12424): all measurements switch to meters, with scaling adapted per model. This fix is not in 2.12.0 (released on June 9) — it will arrive in a future version (likely 2.13.0).

We’ll have to wait for the next version of Zigbee2mqtt to get the fix

[pierre-gilles]

To facilitate debugging of the AI agent, I will add a button in the Artificial Intelligence integration to download the entire context sent to the agent, as well as the last 10 messages of the conversation.

LLMs are by nature « black boxes », and it is often difficult to reproduce the inconsistent behaviors reported by users. Responses heavily depend on the context of each home, connected devices, configured scenarios, etc. Two users may therefore get very different results with the same request.

@mutmut, this will notably allow me to analyze the problems you encounter more easily. Without access to the exact context sent to the model, it’s quite complicated to understand what’s going on

This will be available in the next version of Gladys!