'Small' here means compute, not memory: 17B of 594B is under 3% active per token
But all 594B still have to sit in memory, about 1.2 TB in BF16 or roughly 300 GB at 4 bits. The Apache 2.0 licence is the other big deal, given today's ERNIE release excludes whole regions. How MoE
My take: OpenAI's most advanced cyber model isn't public
Presumably because offensive capability at that level warrants access controls. DeepSeek and GLM performed close behind it in a simulated network intrusion experiment. Both are free to download. That gap is the story.
This is my big fat fish. I attached a live2d body to her, and then gave her a set of free voice output software.
Someone has reverse engineered Wispr Flow and come out with a FREE alternative that works just as well as the paid version.
It's called Phonon: fast, local, open source voice dictation for Mac. Trained a 1.7B model (LFM3-1.7B) with DeepSeek and the result is within
You have reached the end of the archive
All of deepseek