DeepSeek Harness, the open Claude-Code rival, gets a desktop app
Free, MIT, plugin-based, running DeepSeek-V4. Plus: Janus runs GGUF on any GPU, and Anthropic warns GLM-5.3's safeguards strip off easily.

Copy markdown
A free, open alternative to Claude Code
DeepSeek Harness is MIT-licensed and built on a plugin-everything architecture (Cordis), with subagents, agent teams, and four workflow modes — Standard, Programmatic Tool Calling, Minimalist, and Creation. It runs DeepSeek-V4 and the cheaper V4-Flash, and you can spin up the full web UI with one command: npx @deepseek-ai/dsh web.
The desktop app — and its catch
Preview builds (0.1.7-rc) are out for macOS (arm64, Apple-notarized) and Windows (x64); no Linux yet. Unlike the free web and CLI paths, the desktop client requires a topped-up account plus real-name verification — so for free, scriptable use, stay on dsh web or the CLI.
Janus: run any GGUF model on any GPU
A new MIT Go binary serves .gguf models over an OpenAI-compatible API with Vulkan acceleration on AMD, Intel, and NVIDIA — no CUDA required. It ships a web UI, built-in tools (file ops, OCR, math), hot model-swapping, and drops into Cursor, Cline, or any OpenAI client. Local inference on the card you already own.
Anthropic flags GLM-5.3's strip-off safeguards
Anthropic published testing that the open-weight GLM-5.3 (Zhipu / Z.ai) can find vulnerabilities and build end-to-end exploits at near-Claude levels, and that "abliteration" cut its refusal rate from ~90% to ~2-3%. Expect the open-weights access debate — and scrutiny on self-hosting frontier open models — to heat up.