DeepSeek: 48 minutes, 4 of 5 milestones.
Qwen: 6.5 hours, 2 of 5. Both fail the pixel-fidelity check against real DOSBox. Neither plays its own game. But if I had to pick a winner, Deepseek V4 Flash took this easily.
Let me give you a feel for the speed. I think it’s very capable. There are about 50~60 tokens. I didn’t test it in detail. I just felt it visually.
I gave him a task last night and it was all automated. It was really awesome. I think there is no need to use GPT for tasks that are not top-notch in the future. Offline DeepSeek works very well.
We no longer have deepseek v4flash but we continue with a model of dubious origin.
Quick racing game build test for deepseek flash vision exp
Build-in vision looks like have better coherent layout - feel not well trained to use it vision ability (maybe harness issues) - feel think a bit less than flash v4 0731
You have reached the end of the archive
All of deepseek