Jin Daily AI Trivia - DeepSeek-V3.2-Exp Dropped. Now everyone can run AI (SO CHEAP!!!)

Jin Daily AI Trivia - DeepSeek-V3.2-Exp Dropped. Now everyone can run AI (SO CHEAP!!!)

DeepSeek-V3.2-Exp is the experimental DeepSeek Sparse Attention (DSA) model that will serve as the foundation for DeepSeek-V4.

According to DeepSeek, its performance is on par with V3.1-Terminus (which also has reasoning mode), but V3.2 slashes the price to rock-bottom levels.

Pricing (1M input tokens + 1M output tokens, no cache hit):

DeepSeek V3.2: $0.70 Gemini 2.5 Pro: $11.25 GPT-5: $11.25

With cache hits:

DeepSeek V3.2: $0.448 Gemini 2.5 Pro: $10.31 GPT-5: $10.125

That’s insane—only ~4.3% the cost of other SOTA models. A total price slaughter.

I study the paper. https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf

They basically reinvent the whole token indexer. (lightning indexer top-K selection, then do a multi head attn)

I’ll be testing the new API soon and will share my findings later.

THIS. IS. MADNESS.