·1 хв читання · 118 слів·👁 14.3K↗ 8
🐳 DeepSeek-V3.2: New Model from Chinese Startup
DeepSeek has released a new experimental V3.2 AI model. It's based on the recent V3.1-Terminus, but with a new DeepSeek Sparse Attention mechanism.
The improved architecture enhances the model's efficiency with long context. It is cheaper while maintaining the same level of intelligence.
Still, in some tests that are sensitive to the amount of "thinking" before an answer, there are small drops in performance. The developers note that these appear due to shorter "reasoning" by the model. And if it spends a similar number of tokens, the gap disappears.
UPD: DeepSeek-V3.2 is now available for free in out bot:
1️⃣ Go to @GPT4Telegrambot
2️⃣ Click Choose Model button
