Blog

Qwen4 Status Check: What's Officially Confirmed, What Hasn't Shipped, and Is There a 35B?

Qwen4 has been officially named, but it is only "in training": no model card, weights, API, pricing or benchmarks. The four-tier naming (Max/Plus/Flash/27B) comes only from on-stage slides at the conference, and the 5–10T parameter figure refers to Qwen4.5/Qwen5. This post separates official confirmation from rumor, covers the only architecture preview you can actually run today, Qwen3.8-Flash-Next (125B MoE, 6B active, GDN+QSA hybrid attention, qwen-community-1.0 license), the engineering-side evidence (qwen4_exp has already landed in transformers/vLLM/SGLang), and answers the 35B question: it was never mentioned officially; 35B belongs to the Qwen3.5/3.6 and UI-Agent lines, and community projects have already ported 35B-A3B onto the Qwen4 architecture. A local VRAM requirements table is included.

AI Rewrites the Adobe Suite and Office: Clean-Room Clones — Progress, Controversy, and the One Change That Really Matters

Starting in late September 2026, a developer used Claude Opus 5.5 with multiple agents in parallel to do a pure-Rust clean-room rewrite, shipping 7 open-source apps that take on the Adobe suite within a week, followed by counterparts to Word/Excel/PowerPoint; PhotoCraft gained 53,000 stars in a week. Using the project's own completion documents, this post unpacks why "it runs ≠ it's usable": 76% breadth, 45% able to do real work, only 3 of 25 performance budgets met, and 1,100–1,800 agent hours still needed to reach beta. It also spells out where the real impact lies — a command registry + CLI + JSON + MCP, which for the first time lets AI drive design, video, and document tools headlessly in batch.

Orange Pi Buying Guide: A 352 TOPS Ascend Compute Card, and Low-Power Boards for Running AI Agents

For anyone who wants to run models locally or build an AI Agent. This post splits Orange Pi into two routes: first, the Ascend compute card (AI Studio Pro: 2×310P, 352 TOPS, 96/192GB, but only 408GB/s of bandwidth, so it "fits, but doesn't run fast"); second, low-power boards for running AI Agents (the Agent itself needs little compute; the real threshold is 2–8GB of RAM; measured idle draw is 1–2W, or 7–58 yuan a year in electricity). Includes tier tables for both routes, a power and electricity cost comparison, a hybrid NPU + cloud API setup, and the pitfalls.

Introducing Tauri: A Framework for Building Cross-Platform Desktop Apps with the System WebView

Tauri is a framework for building desktop and mobile apps with a web frontend and a Rust backend: instead of bundling a browser engine, it uses the WebView that ships with the operating system, so installers are typically just a few MB. This article covers its architecture, how it differs from Electron, what it can and can't do, its security model, and the current state of its releases.