US companies are not using Claude or Codex.
Too expensive. They run on Quen, DeepSeek, and Kimi instead. My move: use Opus as the orchestrator, then let it delegate work to cheaper models. Cheap tokens, same output.
I don't know if this will be helpful to anyone, but I did something quite special when creating #DeepSeek-chan this time.
By making multiple references to both the original image and the see-through image, we succeeded in having GPT generate a single part with high image quality while inferring the back of the part. Well, there are some mistakes.
The first build of Qwen3.8-Flash IQ3_XXS had bugs.
DeepSeek V4.1 Flash tried 5 times 20 minustes and couldn't fix them—but Qwen3.8-Flash IQ3_XXS fixed them in the second build. For coding, Qwen is definitely better.
تصدقون إن HarnessRouter فتحوا السورس حق طبقة بنية تحتية plug-and-play تشغّل Codex و Claude Code و…
تصدقون إن HarnessRouter فتحوا السورس حق طبقة بنية تحتية plug-and-play تشغّل Codex و Claude Code و Hermes و DeepSeek Harness و System One powered by Jev و9 غيرها من harnesses الوكلاء تحت واجهة وحدة؟ 😳 مع Unified Harness Protocol و API متوافق مع OpenAI Responses. ← تجيب Jev…
You have reached the end of the archive
All of deepseek