SITUASI DIJELASKAN: DeepSeek V4.1 Flash mengalahkan model Pro-nya sendiri dari 3 minggu lalu.
552B campuran pakar, 8B aktif untuk input dan 16B untuk output. Berlisensi MIT, konteks 1 juta • Mengalahkan V4 Pro di semua benchmark kecuali GPQA Diamond dan Ujian Terakhir Kemanusiaan • masukan $0,15, $0,60
Claude Opus 5 charged 20x more than DeepSeek V4.1 Flash and delivered a white screen.
$0.59 vs $0.03. Same prompt. Opus 5 took longer and rendered a sun so blown out you cannot see the ocean. DeepSeek rendered an actual sunset for 3 cents. DeepSeek V4.1 Flash is hit or miss. I
Inspired by developedbyed's AssBench, I decided to try AssBench on Deepseek V4.1 Flash.
Needless to say, this isn't as good as what Astra did. It also took Deepseek like 8 hours to finish this. Waited for a whole day for this.
It's interesting that the new DeepSeek 4.1 flash is also using very large ngram embeddings ( can be…
It's interesting that the new DeepSeek 4.1 flash is also using very large ngram embeddings ( can be stored on cpu ram ) similar to the upcomming Qwen4 architechture. Here's a fully 3D breakdown of every weight on the model and what it does. Enjoy.
Sudah sampai ujung arsip
Semua deepseek