DeepSeek V3.2 is an in-place architectural revision of DeepSeek V3, keeping the same 671B sparse MoE frame (~37B active parameters) while trimming inference cost by roughly 30% versus its predecessor. If you're routing production traffic through V3 today, V3.2 is not a lateral swap — it is the same model family with meaningfully lower per-token rates at most providers. Both models use Multi-Head Latent Attention and DeepSeek's FP8 mixed-precision training pipeline. The difference is mostly in post-training: V3.2 received additional instruction-following refinement and updated RLHF data, which narrows the gap to frontier reasoning models on coding and math benchmarks without inflating active-parameter count. [DeepSeek V3](/models/deepseek--deepseek-v3) holds a pricing advantage on providers that have not yet migrated their weights — you may find lower spot rates on burst capacity tied to the older checkpoint. For batch inference jobs where latency is unconstrained, the rate difference can offset V3.2's benchmark gains and make V3 the cheaper option per correct output. [DeepSeek V3.2](/models/deepseek--deepseek-v3.2) wins on interactive workloads: the additional instruction-following tuning reduces refusals and format drift on long system prompts, making it more reliable for structured-output agents and multi-turn chat pipelines. Pick V3 if your workload is batch-oriented, your provider still offers a lower per-token rate on the old checkpoint, and benchmark delta is immaterial to your use case. Pick V3.2 if you're running interactive agents, structured-output pipelines, or any task sensitive to instruction-following fidelity — the cost reduction over V3 makes it the default upgrade path.