Highlights
Web search reliability overhaul. Ka1zen stopped confidently telling users that real, recently-released products "don't exist". Search now fires when it should, runs a reformulated query (not your raw conversational sentence), the model can no longer fabricate a search in its head or dismiss live results as "simulated", blocked pages fall back to a real search, and vision models participate too. The result: relevant, grounded, sourced answers, verified end-to-end.
DFlash speculative decoding for code (GGUF). A new optional "DFlash for code" toggle (Settings → Performance) roughly doubles code generation on the slow dense models, lossless — Qwen 3.6 27B ×2.63, Gemma 4 31B ×1.91. Off by default; MTP stays the everyday Fast Mode.
Pinned llama.cpp bumped b9585 → b9859 (enables DFlash; MTP non-regression verified bit-identical).
Plus: the first-run wizard no longer pre-enables experimental Fast Mode, response bubbles fill the window width, and a few smaller refinements. See the full changelog.
Install
Download Ka1zen-0.5.0.dmg, drag Ka1zen into Applications. First launch runs a quick setup (Python packages + a pinned llama.cpp build). Requires macOS 15+ on Apple Silicon.