[Verification] You can see 9 years of evolution with your own eyes without reading the paper 🔥
2017 Transformer paper and the latest DeepSeek V4.1 Flash read the paper, A site where structures are arranged in 3D is a hot topic. ・On the left is the original paper from 2017, 6 layers + 6 layers, The one on the right is the 2026 version, 20 layers + 20 layers
OMG, I gave the same one-shot prompt to 4 different models and the outputs are WILD
One of the models surprised me the most while timeline is busy malding over $200 subs for fable 5.1 and astra, i benchmarked the cheap tier: deepseek v4.1 flash, muse spark 1.3, gemini 3.8
DeepSeek-V4.1-Flash、M2 Ultra 192GB + DwarfStar Q2 で 15 tok/s。
You have reached the end of the archive
All of deepseek