Jan v0.8.6 Released with Fixes for Linux, Turing GPUs, and MTP

Jan v0.8.6 Released with Fixes for Linux, Turing GPUs, and MTP

At a Glance

Item Value
Repository janhq/jan
Version v0.8.6
Published 2026-10-09
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 “v0.8.6" of “janhq/jan“, an open-source ChatGPT alternative application that runs 100% offline on PCs, has been released. janhq/jan is primarily developed in Rust and is used by many engineers as a tool to easily run LLMs in local environments.

This update is a hotfix to resolve critical bugs in the previous version, v0.8.5. Particularly for users utilizing Linux (AppImage version) or environments equipped with older NVIDIA GPUs, it is an important update that resolves force-quits and errors when loading models. If you are using v0.8.5, it is recommended to update to this version promptly.

Key Changes

Fixed Crash Issue in Linux AppImage Version

An issue where the AppImage version of Jan would force-quit (crash) every time a model was loaded on certain Linux distributions has been fixed.

This issue was caused by the build machine’s Vulkan loader being mistakenly bundled into the AppImage. Starting with this version, it has been fixed to correctly use the loader installed on the user’s system. This fix affects all users utilizing the AppImage version in Linux environments, and after updating, models can be loaded stably without changing any settings.

Fixed Startup Errors on Older NVIDIA Turing GPUs

An issue where Windows and Linux CUDA engines failed to work on environments equipped with GPUs adopting NVIDIA’s Turing architecture (such as the GTX 16xx series and RTX 20xx series) using graphics drivers older than version R595 has been fixed.

Previously, an error stating “the provided PTX was compiled with an unsupported toolchain" occurred, but with this update, native Turing code is bundled again, similar to the earlier v0.8.4. This allows users with affected GPUs to run Jan as is, without needing to update their drivers to the latest version.

Fixed Severe Speed Drops with Qwen3.8 (Featuring MTP)

An issue where the inference speed of Qwen3.8 with MTP (speculative decoding) enabled dropped significantly after updating to v0.8.5 has been fixed.

Previously, when the “Parallel Sequences" setting was set to “auto", the engine forcibly allocated memory for 4 slots. Because of this, in environments using video cards with around 12GB of VRAM, model data was pushed out to main memory (shared system memory), significantly lowering processing speed.

With this fix, when running MTP-equipped models, the number of allocated slots is kept to 1 if Parallel Sequences is set to “auto". Additionally, values manually set by users are respected as-is. Users running MTP models like Qwen3.8 in ~12GB VRAM environments can recover their original inference performance by updating.

Fixed Image Processing Bugs Returned from MCP Tools

An issue where image data returned by MCP (Model Context Protocol) tools (such as read_media_file which reads media files from the file system) was not correctly sent to the model as an “image" has been resolved.

Previously, image data was sent to the model as-is in the form of massive base64-format text. As a result, not only could the model not recognize the image, but the transmitted data size also exceeded the provider’s input limits, causing a “Provider returned error" message to appear on screen and breaking functionality.

In this version, image data is sent to the model in the appropriate format. Furthermore, for models that do not support image recognition, short annotation text is sent in place of images. Additionally, an issue has been resolved where user questions were mistakenly deleted because the character count of image data was counted as-is during conversation history trimming associated with context limits. For users building workflows involving image inputs utilizing MCP tools, this is an update that significantly improves reliability.

Added Link to Methods for Using Different Llama.cpp Versions

A link to official documentation explaining how to use a version of llama.cpp different from the bundled one has been added under “Settings" -> “Model Providers" -> “Llama.cpp" in the settings screen.

This is achieved by launching llama-server yourself and registering it as a custom provider. It can be utilized as an effective workaround when compatibility issues arise between the bundled inference engine and specific models.

Specifications

In this version, important improvements and compatibility recoveries have been made regarding specific GPU architectures and resource allocation during model execution.

NVIDIA Turing Generation GPUs (GTX 16xx / RTX 20xx Series)

In environments equipped with GPUs adopting NVIDIA’s Turing architecture (such as the GeForce GTX 16xx series and RTX 20xx series), CUDA engines for Windows and Linux can once again be used normally. Even when using graphics drivers older than version R595, native Turing code is provided just like in v0.8.4, allowing model inference processing to be executed without updating drivers.

Linux Environments (AppImage) and Vulkan

Model execution using the AppImage package has been stabilized across general Linux distributions. The issue where the build environment’s Vulkan loader was mixed in has been resolved, and the system’s Vulkan loader introduced in the host environment is now referenced directly, allowing models to be loaded and run using the Vulkan backend without crashing.

Qwen3.8 (Models with MTP / Speculative Decoding)

Execution efficiency for Qwen3.8 models equipped with MTP (Multi-Token Prediction / speculative decoding) has been improved. When the “Parallel Sequences" setting is set to “auto", the number of reserved memory slots has been revised from the conventional 4 slots down to 1 slot. This prevents models from being swapped out to shared system memory (main RAM) even on graphics cards equipped with around 12GB of VRAM, enabling high-speed inference utilizing the GPU’s inherent speed.

Vision-Supported and Non-Vision Models (Image Input via MCP)

In MCP tool integration, image inputs to image-capable models (vision models) are officially supported. Furthermore, even for text-only models that do not support image recognition, short annotation text is sent instead of image data, allowing conversations to continue without causing errors.

How to Get It

Jan can be updated by obtaining the latest installer from the official website or GitHub distribution page, or by using package managers for each platform.

Direct Downloads for Each Platform

Download the installer matching your operating system and run the installation.

  • Windows Environment: Download and run jan.exe.
  • macOS Environment: Download jan.dmg (Universal format) and place it in the Applications folder.
  • Linux Environment (deb package): For Debian or Ubuntu-based environments, download jan.deb and install it via your package manager.
  • Linux Environment (AppImage): As a distribution-independent format, download jan.AppImage, grant it execution permissions, and launch it.

Also, separate installation instructions are provided for Linux Arm64 environments. Check the Linux Arm64 Explanation Page for detailed information.

Updates from App Stores and Package Distribution Services

If installed via a store, you can use the update features of each service.

The installer itself and the complete changelog can also be checked from the Official Website (jan.ai) or GitHub Releases.

Related Articles

What to Read Next

Sources