A Technical Tour of the DeepSeek Models from V3 to V3.2
Source
Sebastian Raschka, PhD
Author
Sebastian Raschka, PhD
Date
Key takeaways · AI-distilled
Multi-head latent attentionThe mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.Full definition → squeezes keys and values into a smaller latent space before they enter the KV cacheThe memory a model keeps about text it has already read, so generating each new token doesn't require reprocessing the whole conversation.Full definition →, then projects them back up when used. You pay an extra matrix multiply to buy a much smaller memory footprint.
The reinforcement-learning ladder, plainly: classic RLHFReinforcement learning from human feedback — training a model to prefer answers humans rate as better, which turns a raw text predictor into a usable assistant.Full definition → needs a reward model and a critic. GRPO drops the critic. RLVR with GRPO drops the reward model too, scoring answers with a compiler or calculator instead of learned human preference.
Hybrid versus dedicated reasoning models keeps flip-flopping. Qwen3 started hybrid then split into separate instruct and reasoning models because each was easier to train and better at its job; DeepSeek went the other way, R1 to hybrid V3.1 and V3.2.
DeepSeek shipped V3.2-Exp months before V3.2 for a practical reason: its sparse attention variant needs custom inference code, so the experimental release existed to get serving infrastructure ready for the real model.
Terms in this piece · Glossary
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
KV cache — The memory a model keeps about text it has already read, so generating each new token doesn't require reprocessing the whole conversation.
RLHF — Reinforcement learning from human feedback — training a model to prefer answers humans rate as better, which turns a raw text predictor into a usable assistant.
Why it matters
If you're evaluating open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → LLMs for agentic work, this walks through DeepSeek's architectural evolution from V3 to V3.2 — including the DeepSeek Sparse Attention mechanism and RLVR reasoning training — so you understand the tradeoffs behind a model competing with GPT-5 and Gemini 3.0 Pro.
Key quotes
“DeepSeek-V3.2 uses exactly the same architecture as DeepSeek-V3.2-Exp”
“For reasoning and agent tasks, we employ rule-based outcome reward, length penalty, and language consistency reward. For general tasks, we employ a generative reward model where each prompt has its own rubrics for evaluation.”
“All experiments used a single model, our final proof generator, which performs both proof generation and verification.”
