anand 
vidya 
bhavesh 
naresh
Why do AI agents re-read the whole conversation every tool call? DeepSeek fixed that.
The problem
Every time you ask an AI agent to do a step (run a tool, read a file), it has to re-read the whole conversation from the start before it can answer. Like re-reading an entire book every time someone asks "what happened on page 200?" That's slow and expensive, and agents do it hundreds of times per task.
What DeepSeek did
They split the model into two halves:
- Reader half (encoder): reads the story once and writes short notes.
- Answer half (decoder): doesn't re-read the story, it just looks at the notes.
The notes are 4× smaller than what other models keep, so it fits more in memory and runs faster (~420 words per second, up to 1 million words of context).
Why the numbers matter
- 552B params = the model is huge (lots of knowledge stored).
- 8B active = but for each word it only "wakes up" a tiny slice of that brain. Cheap to run, smart like a big model.
Why it's helpful
Agents that call tools in a loop (coding agents, research agents) spend most of their money re-reading history. This makes that re-reading roughly half the cost and the memory a quarter. So: longer tasks, cheaper, faster.
Source Link: zartbot.github.io/blog/model_arch/dsv41…