Release Notes

Navigator Release Notes

• Release v2.246.0

  Release Date: April 9, 2025

  Summary

  Performance improvements to webFrame, and support for additional Deepseek V3 and R1 quantizations. Organizations and Remote Deployments available for select customers/prospects/users. Bug fixes and additional improvements.

  New Features

  webFrame Fixes and Additional Model Support

  Better speed performance and stability from models run on webFrame. New quantizations of Deepseek V3 and R1 available in Navigator and webFrame.

  Technical Details:

  Support for Deepseek quantization:

  • Deepseek R1 4 bit
  • Deepseek R1 Distill Llama 8B
  • Deepseek V3 4 bit

  User Impact:

  • Available in the Model Dropdown of the LLM Element
  • 4 bit quantization for V3 and R1 means users can now run either on a single mac studio or 3 mac minis (all high RAM specs, not base)
  • Deepseek R1 Distill Llama 8B can use our optimizer and needs less than 15GB of RAM

  Organizations and Remote Connectivity

  Create a team of users and share access to your clusters' resources and deployments.

  Technical Details:

  • Runtime has added a Homeport Database that allows for managing users and remote connectivity to users' Runtimes
  • Can be deployed to a local machine or to the cloud, depending on what the user wants
  • Current implementation is webAI's Homeport DB for testing purposes, but customers can't deploy their own yet
  • Teams who want to test this will need to be in sales process or current customers who want to test the alpha
  • Can only add users to teams within that Homeport DB, no connections across multiple organizations

  User Impact:

  • Users can invite anyone to their organization, and join as many organizations as they like
  • Allows users to interact with models deployed by their team regardless of what machines they are running on
  • Users can deploy flows to each other's clusters
  • Users can manage the compute resources and deployments in each other's clusters

  Improvements

  Model Changes When Using Groq API Key

  Mixtral 8x-7b-32768 was deprecated in Groq, so we replaced it with Qwen 32B for dataset generation.

  User Impact:

  • More advanced model used in Dataset Generation for Groq API Key users

  Bug Fixes

  MLX Driven Wired Memory Limits Are No Longer Enforced

  Issue: On Mac Sequoia v15.3.1, we were not able to run Deepseek v3 (352 GB) even though we had 512 GB RAM available

  Resolution:

  • This is now handled as part of MLX and does not require user intervention in the terminal

  Previously required manual commands:

  sudo sysctl iogpu.wired_limit_mb=471040

  sudo sysctl iogpu.wired_lwm_mb=471040

  User Impact: No user intervention required in the terminal

  Python Process Cleanup on Navigator Close

  Issue: Python processes related to stopped canvases would sometimes stay active after a user closes Navigator by clicking the red X-button

  Resolution: Process now cleans itself up and Navigator shuts down as expected

  User Impact: No more unexpected background Navigator processes running

Companion Release Notes

• Coming Soon