Where to read the DeepSeek-V3.2 paper
The paper sought by “DeepSeek 3.2 paper” is DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models, a technical report by DeepSeek-AI and its coauthors. It was submitted to arXiv on December 2, 2025, as arXiv:2512.02556. Read the arXiv record or the paper PDF in DeepSeek’s Hugging Face repository. This is the V3.2 report, distinct from the earlier V3.2-Exp report. (arxiv.org)
What the report’s evaluations say
The report’s abstract characterizes DeepSeek-V3.2 as performing comparably to GPT-5 on the authors’ reasoning evaluation, and says the higher-compute V3.2-Speciale surpasses GPT-5 and reaches reasoning performance on par with Gemini 3.0 Pro. These are the authors’ reported comparisons, not independent validation. The report’s benchmark figure notes two scope details: its HLE results use the text-only subset, and its HMMT 2025 results use the February competition to match the baselines. The available report text here does not give the relevant scores or full evaluation protocols, so these comparisons are a high-level signal—not enough to infer a likely score on your own tasks. (arxiv.org)
For coding-agent selection, the report also says V3.2 significantly outperforms open-source language models on SWE-bench Verified and Terminal Bench 2.0. That supports considering it for further evaluation on coding workflows; it does not establish how it compares with every closed model or how it will perform in a particular harness, budget, or production environment. The report describes DeepSeek Sparse Attention and a large-scale agentic-task synthesis pipeline as parts of its approach. DeepSeek’s release announcement specifies that the synthesis covered more than 1,800 environments and 85,000 complex instructions. Those figures describe the reported training-data pipeline, not a guarantee of task success in deployment. (arxiv.org)
Production API choice: a tool-using assistant
The paper can inform model selection, but it does not establish which model identifier is currently served on a hosted API route. In the DeepSeek API models-and-pricing documentation checked on November 10, 2026, deepseek-flash maps to DeepSeek-V4.1-Flash-1017. The same table marks both thinking mode and tool calls as supported for that model. The current table does not list V3.2; that omission does not, by itself, prove that no V3.2 route exists, so confirm route availability directly before depending on it. (api-docs.deepseek.com)
If a current documented model is acceptable instead of V3.2, deepseek-flash is a documented candidate to evaluate for a tool-using assistant. Its listed rates per million tokens are $0.30 for uncached input, $0.006 for cache-hit input, and $1.20 for output at peak; off-peak rates are $0.15, $0.003, and $0.60, respectively. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays; the documentation says prices may change. These are API billing details, not a claim that Flash will match V3.2’s research results. (api-docs.deepseek.com)
Selection takeaway: use the V3.2 report as research evidence for its stated reasoning and coding-agent evaluations, while treating those results as author-reported and benchmark-specific. For a production tool-using workload, the current documentation supports evaluating deepseek-flash for its listed model version and capabilities. If V3.2 itself is required, verify the exact serving route, capabilities, and terms rather than assuming the paper or a past launch announcement establishes present API availability.
Sources
DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models