Skip to main content
Triton AI Triton AI
  1. TritonAI
  2. TritonGPT
  3. Release Notes
  4. September 8, 2026 Release Details

September 8, 2026 Release Details

Overview

This release upgrades TritonGPT to version 4.6.6 in production. It introduces a new default model, refreshes the model lineup, and adds Voice Mode, a coding sandbox, and adjustable reasoning effort. The configuration was tested on the stage environment before deployment, and the UC San Diego Assistant remained available in a limited form during the change window.

What’s New and Changed

New default model

Muse Glimmer 30B, an open model hosted on UC San Diego infrastructure, is now the default model for TritonGPT assistants. It replaces the previous default and runs more efficiently, improving response speed and adding capacity during periods of peak demand, including course scheduling at the start of the term.

Refreshed model lineup

  • GPT-5.6 Terra replaces GPT 5.4 as the OpenAI flagship option.
  • Gemini 3.7 Flash replaces Gemini 3.5 Flash for fast responses.
  • Claude Sonnet 5 replaces Claude Sonnet 4.6.

The available catalog changes as models are evaluated, added, updated, or retired. Use the Model Hub for current availability, capabilities, and published rates.

Voice Mode

Voice Mode lets you speak with a TritonGPT assistant and hear its responses, in addition to typing. It is available alongside the existing chat interface and is now enabled for all assistants.

Click the microphone icon to start. Your speech is transcribed as you talk, and responses are read back as audio with the full text displayed alongside. Voice Mode requires microphone access in your browser, and you can configure auto-send, auto-playback, and playback speed in user settings.

Coding sandbox

The coding sandbox is enabled for all assistants. It lets assistants write and run Python code to answer questions that need calculation or data analysis, with output returned in the conversation. Assistants run code automatically when a question calls for it.

What this makes possible:

  • Uploaded files, including spreadsheets and CSVs, can be analyzed in chat.
  • The sandbox can create and return files, such as reports and processed datasets.
  • Charts and graphs render directly in the conversation.
  • Large or complex calculations run as code and return exact results.

Common Python data libraries, including pandas, NumPy, SciPy, and matplotlib, are available. Code runs in an isolated environment with no network access and no access to campus systems, and each run has a time limit.

Adjustable reasoning effort

You can now adjust how much reasoning a model applies to a response. Choose a lower setting for faster answers or a higher setting for work that needs more depth. The control is available per model in the chat interface.