The technical report attributes V3.2's advantages to sparse attention, scalable RL, and large-scale Agent task synthesis. The goal is to reduce costs and improve tool generalization in long contexts, but the report's benchmark and API production performance still require independent reproduction.
Models: DeepSeek-V3.2, V3.2-Exp, and V3.2-Speciale; the report covers standard reasoning, long-context, Agent, and human-preference evaluations.
Architecture: DSA (lightning indexer + fine-grained top-k selection), followed by specialist distillation and mixed RL after continued pretraining.
Training/evaluation data: The Agent synthesis pipeline covers 1,800+ environments and 85,000+ complex instructions; the report also mentions 128K long-context training and H800 inference cost estimates.
The report's detailed benchmark tables/figures are not fully available as reproducible text. Its method description includes specialized thinking/non-thinking tracks, six domains (mathematics, programming, logic, general Agent, Agentic coding, and search), and training factors such as outcome reward, length penalty, and language consistency reward.
The report's abstract states that V3.2 is close to GPT-5 on multiple reasoning benchmarks, while V3.2-Speciale surpasses GPT-5 and approaches Gemini-3.0-Pro; Speciale achieved gold-medal performance at IMO/IOI 2025.
DSA reduces the main attention complexity from O(L²) to O(Lk), with each query selecting 2,048 key-value tokens (the report's sparse-training configuration).
Continued pretraining's sparse stage: 1,500 steps, 480×128K sequences per step, and approximately 943.7B tokens; these are training settings, not parameters required for deployment.
The report says V3.2-Exp scored 4 points higher than V3.1-Terminus on Artificial Analysis Long Context Reasoning; it also cites multiple areas of leadership on Fiction.liveBench, but these are external evaluations cited by the report, not independent reproductions in this review.
V3.2's architecture and training approach target long-context efficiency and Agent generalization, making it suitable as an open-weight Agent baseline for research. Actual services should nevertheless be validated separately across short/long contexts, thinking/tool use, and provider parsers.
This is a self-reported DeepSeek technical report. Benchmark selection, prompts, scaffolding, and evaluation details are incomplete; conclusions about rankings should not be based on the abstract alone.
Training token counts, top-k, and DSA complexity do not directly determine API costs or user-visible quality.
The V3.2, V3.2-Exp, and Speciale checkpoints and capabilities differ; comparisons across variants can be confounded.
Artificial Analysis and Fiction.liveBench results should be checked against their original pages.
Select the model variant, weights/service, and task split; fix thinking, tools, context length, and output budget.
Run standard reasoning, coding, long-context, tool-calling, and Agent-outcome tasks in separate layers.
For long contexts, record tokens/s, time to first token, GPU/cost, and accuracy; for Agents, save complete tool traces.
Compare GPT and other open models using the same harness, distinguishing figures from the official report, external results, and your own reproduction results.
DeepSeek V3.2
