Jan v0.8.8 Released: llama.cpp Upgrade and Security Enhancements

Jan v0.8.8 Released: llama.cpp Upgrade and Security Enhancements

At a Glance

Item Value
Repository janhq/jan
Version v0.8.8
Published 2026-10-28
License NOASSERTION
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

The latest version of Jan, v0.8.8—an open-source alternative to ChatGPT that operates 100% offline in a local environment—has been released. This update is a security update following v0.8.7, and users are encouraged to update. The most significant change is the major upgrade of the built-in inference engine, llama.cpp, from 0.5.0 to 0.7.0 (b11669).

Breaking Changes / Deprecations

Due to security enhancements, this update introduces several important changes regarding API behavior and settings.

Item Old Version New Version
CORS Settings Setting existed but had no effect Settings are now effective and set to “ON" by default after the update
API Access No Origin restrictions Only trusted hosts (same machine, Jan itself, or hosts registered in “Trusted Hosts") are permitted
Data Folder Movement No restrictions Moving to system folders, the home folder itself, or the root of a drive (e.g., D:\) is denied

If you are calling Jan’s API from a browser-based client (such as Open WebUI), you must add that host to the new “Trusted Hosts" settings. Additionally, users who have placed their data storage location directly at the root of a drive will need to change it to a subfolder, such as D:\Jan.

Key Changes

Upgrade to llama.cpp 0.7.0

The built-in engine has been updated to llama.cpp 0.7.0 (b11669). This enables new model-specific settings, such as “MoE expert GPU cache" for Mixture-of-Experts (MoE) models, and “Probabilistic draft sampling" and “Draft batch size" for speculative decoding. Since the engine’s default values are applied, there will be no change in behavior unless these are explicitly configured. Note that prompt caches saved with the previous engine will be discarded and rebuilt.

Security and Privacy Enhancements

Links within chats will now open in the system’s browser rather than within the Jan window. Additionally, agent file tools are now restricted from following links outside the workspace, and restrictions have been strengthened to prevent home directories or credential folders from being attached as writable folders. Furthermore, integration with Claude Code has been improved to ensure safer writing to shell profiles.

Bug Fixes for Windows Environments

A problem where local model loading failed on Windows when the username contained non-English characters (such as Cyrillic) has been fixed. Additionally, improvements have been made to ensure system tools are executed from Windows system folders on Windows.

Improved Behavior with Proxy Settings

A problem where processing documents with “Use embeddings" selected failed in environments with an HTTP proxy configured has been fixed. Requests to the Jan engine are now made directly without going through the proxy; the proxy will continue to be used only when utilizing remote providers.

MLX Server Management on macOS

A problem on macOS where the MLX server remained in memory even after Jan was closed has been fixed. The server will now stop when Jan exits, and if a server from a previous session remains at startup, it will also be stopped. However, this does not affect MLX servers started manually by the user.

Supported Models and Hardware

Due to the engine upgrade, “MoE expert GPU cache" settings are now available for Mixture-of-Experts (MoE) models. Additionally, when using speculative decoding, the new parameters “Probabilistic draft sampling" and “Draft batch size" are available.

How to Get It

You can update Jan by downloading the installer for your respective OS. Please download the file suitable for your environment from the official website or the GitHub release page.

For more details, refer to jan.ai or GitHub Releases.

Related Articles

What to Read Next

Sources