After releasing the open source version of DeepSeek Harness Desktop yesterday, we quickly iterated on 7 minor versions.
🔥 Latest online: Preset Plaza. Now you can directly browse different Agents and import the corresponding Presets and Skills with one click, eliminating the need to manually copy files and configure configurations. With one click, Agent can quickly have a complete workflow.
Dropped the $200/month Claude subscription.
Swapped Opus 5 for Deepseek V4 Flash at roughly 1/20th the cost. Only snag: it’s blind—no images, no video. Then I stumbled on a two-install hack that bolts vision onto basically any text-only model. Here’s the play: 1/ Install
This "The Japan prompt" also sounds pretty good.
DeepSeek V4 Flash 0731, I feel like the bugs have decreased considerably since I changed the harness to DeepSeek Harness.
DeepSeek V4 Flash is supposed to be the smaller, faster model.
Instead, it beat DeepSeek V4 Pro across all 9 agent benchmarks. I tested it across 50+ builds. Here's what stood out:
You have reached the end of the archive
All of deepseek