DeepSeek's Creator Mode turns plain English into Harness plugins
Seven copyable prompts ship with it. Plus three open-weight drops to pull today: Deep Principle's 753B science model, Nace AI's 27B, SenseTime's 9B.
Copy markdown
Describe the plugin, Harness builds it
DeepSeek showed off Creator Mode for Harness, its open-source agent platform: describe a tool in plain English and it generates a working plugin — demos span a video editor, a paper reader, interactive 3D visualizations, a poster designer and a stats dashboard. The seven demonstration prompts are published in English and Chinese, so you can paste them into Harness, pick Creator Mode, and build your own version today.
A 753B open-weight model that out-sciences GPT-6-Astra
Deep Principle open-sourced Hydrogen-1.0, a ~753B MoE post-trained from GLM-5.3 and built for scientific work — reaction prediction, NMR structure elucidation, lab-protocol compilation and materials design. On its own evals it leads GPT-6-Astra by 5.13 points on scientific computing and 4.76 on forward reactions; weights are on Hugging Face. Worth grabbing if you build research or chemistry tooling — just budget serious GPU for a model this size.
Nace AI's 27B decision model runs on your Mac
Drex Max v1 is a 27B open-weight multimodal 'decision' model with a 262K-token context, released under Open RAIL-M and small enough to run locally on Apple Silicon. It scores 67.52 on Decision Index 0.4 (+6.27 over Drex v2.0) and 97.3 on InfographicVQA — a weekend grab for agents that read dashboards and documents and then act on them.
SenseTime's 9B looped multimodal, Apache 2.0
SenseNova-Loop-9B-Preview is an Apache-2.0 multimodal model (text and image in, text out) built for long-context, agentic workflows — document understanding, coding and tool use — trained to 262K tokens. Weights are on Hugging Face with a single-GPU example plus llama.cpp and LM Studio quants, so it fits on one card.