Preliminary version of DeepSeek Harness available as Electron packages for Windows and macOS
Read more
Pandaily
pandaily.com

Preliminary version of DeepSeek Harness available as Electron packages for Windows and macOS

DeepSeek Harness now has a preliminary version for desktop computers: downloadable Electron installers for Windows and macOS operating systems are available on the website download.deepseek.com, according to reports from 36Kr and AI Front dated around September 25, 2026. The official product name is DeepSeek / DeepSeek Harness.

Previously, Pandaily covered the launch of the open-source Harness Agent runtime and the Tencent WeKnora plugin; this review focuses specifically on the desktop client interface, not a repetition of information about the open framework.

Community builds align with the official update host specified in the deepseek-ai/deepseek-harness repository. Examples of available packages include an executable for Windows x64 and a DMG for Mac arm64 in the 0.1.7-rc.1 branch, which contains a built-in updater showing version 0.1.7-rc.2. Apple certification involves Hangzhou DeepSeek Artificial Intelligence Co., Ltd. with package ID com.deepseek.dsh. At the time of publication of these materials, the packages had not yet been placed on the main DeepSeek page or in the GitHub Releases section, so they should be considered available via a preview or nightly channel, rather than as a general public release. The reports did not contain information about a build for Linux desktop systems.

Technical notes indicate that the desktop application shell utilizes existing Web UI components, as well as agent runtimes, sessions, tools, and plugins, instead of completely rewriting Harness from scratch. The first login may require real name verification, balance top-up, or quota usage, or users can insert an official API key. Product modes cover the categories 'Office and Creation' and 'Code and Development', offering process visibility options (Results Focus, Key Details, Full Process) and agent workflow levels, including Standard, PTC (Tool invocation), Minimalist, and Creation for Cordis plugin debugging.

For developers who already use Harness from the source code, the main news is the appearance of an official Electron preliminary version on the CDN for Win/Mac—this is a weekly package, not a declared product for widespread use or the release of a new base model.

Similar stories

MiniMax integrates M3.1-Flash-Preview model for coding into MiniMax Code product
Read more
pandaily.com

MiniMax integrates M3.1-Flash-Preview model for coding into MiniMax Code product

MiniMax announced on September 27, 2026, that the M3.1-Flash-Preview model is now available within its product for assisting with code writing in MiniMax Code. This information was confirmed by sources, including IT Home, Startup Fortune, and other developer reviews. The official English name of the product is MiniMax / MiniMax Code.

The main focus of this review is on the in-product coding model and the reasoning controls. However, MiniMax has not yet published a full public specification of parameters or an open API for M3.1-Flash-Preview, so these details cannot be confirmed.

The company's marketing message presents this Preview model as an ultra-fast text model designed for daily development. It is suitable for tasks such as debugging, generating function snippets, handling edge cases, regression testing, and checking the impact of changes, rather than being a new flagship for general chat.

Reviews describe a closed loop operation within MiniMax Code that covers problem localization and implementation, as well as verification through testing and subsequent delivery. Developers who noticed the model in the product list a day before the official announcement were already considering it as an SKU for coding workloads, coexisting with older M3 and M2.7 variants in one interface.

A practical difference noted in the English product reports is the five-level reasoning resource slider, which ranges from low to a new maximum level. This slider allows users to balance computation depth against latency when performing routine or more complex coding tasks. This control is located inside MiniMax Code; the Preview endpoint is described as limited to the company's own tool, not as a widely documented public API in this launch.

Parallel promotions include a Token Plan reset and a verification campaign from September 28 to October 7, which offers double points applicable to code models, including M3.1-Flash-Preview. Users should remember that the M3 architecture and SWE-bench metrics relate to the previous M3 flagship, not automatically to Flash-Preview, until MiniMax releases a specific card.

For coding tool users, the main news is that MiniMax Code has received the Flash-Preview coding model with an explicit reasoning effort slider—a component built into the development environment (IDE), not an open-weights release.

Popular