DeepSeek Harness is a small test of its skills. I simply made a dynamic image. It feels very good.
I ran DeepSeek V4 Flash vs Claude Opus 5 on an autonomous Three.js driving challenge
Both runs were routed through Constellation Gate AI _GateAI, which gave me quick and easy access to both models, captured usage, timing and cost across providers. Fresh sessions using
Nobody is talking about the model quietly bleeding out in dead last
Moon dev: "six of the smartest AIs on earth got a hundred bucks each and one of them is already burning through it faster than the rest" "DeepSeek is sitting at the very bottom and what it did on the last hour
When I tried to create a showdown with Gpt5.6 luna and Qwen3.8 27b using the exact same prompt with…
When I tried to create a showdown with Gpt5.6 luna and Qwen3.8 27b using the exact same prompt with deepseek v4 flash, the results suddenly came out on a completely different level! 🙀 This time, after playing the same prompt three times after completion, I could gradually see the quality improve. There was a method like this!
You have reached the end of the archive
All of deepseek