DeepSeek-V3.2-Exp · 官方评测披露

DeepSeek-V3.2-Exp 当前收录 17 项测试的官方记录。已完成列明材料的核查。

模型发布 · DeepSeek V3

官方材料与核查范围

All listed official model-card, official News/API, metadata and technical-report materials were checked through their relevant scope. The later V3.2 report Section 4.3 and Table 5 explicitly attribute this synthetic-agent ablation to V3.2-Exp; its disclosure date is later than the Exp release. Tables without an explicit unit retain an empty unit; figures without exact labels were not estimated. 核对时间 。

当前资料不足以还原首发。 No reliable historical snapshot proving the complete launch-day material set was obtained. Official publication dates and current materials do not substitute for a launch snapshot.

  • DeepSeek-V3.2-Exp official Hugging Face model card · 主文 · HF revision 194c67e12b1b0d6df0ef373ddcf215bc84027409; current card snapshot
    公布 未确认 · 采集 · 已核查 Official model-card Markdown read in full; benchmark rows, notes, configuration statements and linked reports checked.
  • DeepSeek-V3.2-Exp official Hugging Face model metadata · 主文 · author=deepseek-ai; repo createdAt=2025-09-29T06:07:26.000Z (repository upload time only); lastModified=2025-11-18T02:39:54.000Z; sha=194c67e12b1b0d6df0ef373ddcf215bc84027409
    公布 未确认 · 采集 · 已核查 Official model API metadata checked for author, repository identity, revision and timestamps; createdAt was never used as release date.
  • DeepSeek-V3.2-Exp 发布,训练推理提效,API 同步降价 · 主文 · Official DeepSeek News index entry dated 2025-09-29; index downloaded and retained locally
    公布 · 采集 · 已核查 The official dated entry or current service table was read for this release scope; aliases were not treated as weight checkpoints.
  • DeepSeek-V3.2-Exp technical report · 附件 · Official V3.2-Exp report: 6/6 PDF pages read; physical PDF page 4 contains Table 1 and does not contain the synthetic-agent Pass@K Table 5.
    公布 未确认 · 采集 · 已核查 Official V3.2-Exp report: all 6 pages read. Physical PDF page 4 contains Table 1, not the synthetic-agent Pass@K Table 5 cited by the former entries.
  • DeepSeek-V3.2 official technical report; later retrospective evidence for V3.2-Exp · 附件 · Official 23-page V3.2 report published with the 2025-12-01 release. Section 4.3 on physical PDF page 15 describes the 50-instance sample; Table 5 on page 16 explicitly names DeepSeek-V3.2-Exp.
    公布 · 采集 · 已核查 Official V3.2 report checked at Section 4.3 on physical PDF page 15 and Table 5 on page 16. The text describes a random sample of 50 general synthesized agentic tasks; the complete table header names DeepSeek-v3.2-Exp and all three Pass@K rows match. The report was released later and is current-view evidence only.

配置

  • DeepSeek-V3.2-Exp reasoning mode without tool use · Rows are under “Reasoning Mode w/o Tool Use”; exact effort is not stated.
  • DeepSeek-V3.2-Exp agentic tool use · Rows are under “Agentic Tool Use”; exact scaffold and effort are not stated in this table.
  • DeepSeek-V3.2-Exp synthetic-agent ablation · Randomly sampled 50 general synthesized agentic tasks; Pass@K protocol as printed in Table 5.
    未知项 This is a training-data ablation, not the standard capability table.

直接前代与后续发布

直接前代 DeepSeek-V3.1-Terminus。 官方关系依据

V3.2-Exp builds upon V3.1-Terminus by introducing DeepSeek Sparse Attention.
9 项披露延续 · 8 项新增 · 3 项本次未见 · 0 项待核查
DeepSeek-V3.1-Terminus 与 DeepSeek-V3.2-Exp 的披露变化
测试已核查状态DeepSeek-V3.1-TerminusDeepSeek-V3.2-Exp
MMLU-Pro · DeepSeek card version unspecified; Exact Match延续

已记录的评测条件有变化

前代记录本次记录
GPQA Diamond · Diamond subset; task snapshot unspecified延续

已记录的评测条件有变化

前代记录本次记录
Humanity's Last Exam · Text-only subset as stated in V3.1 and later comparison cards; task revision unspecified延续

已记录的评测条件有变化

前代记录本次记录
LiveCodeBench · Version unspecified; Pass@1延续

已记录的评测条件有变化

前代记录本次记录
AIME · 2025 edition; Pass@1新增—本次记录
HMMT · 2025 edition; reported score, metric not stated in V3.2-Exp card新增—本次记录
Codeforces · Rating; contest snapshot unspecified延续

已记录的评测条件有变化

前代记录本次记录
Aider-Polyglot · Version unspecified; Accuracy延续

已记录的评测条件有变化

前代记录本次记录
BrowseComp · Version unspecified; Accuracy延续

已记录的评测条件有变化

前代记录本次记录
BrowseComp-zh · Version unspecified; Accuracy延续

已记录的评测条件有变化

前代记录本次记录
SimpleQA · Internal search agent protocol; task version unspecified延续

已记录的评测条件有变化

前代记录本次记录
SWE Verified · V3.2-Exp comparison; agent/task version unspecified新增—本次记录
SWE-bench Multilingual · V3.2-Exp comparison; agent/task version unspecified新增—本次记录
Terminal-Bench · Task version unspecified in V3.1-Terminus/Exp card新增—本次记录
General synthesized agentic tasks · 50-task ablation; Pass@1; unit printed as %新增—本次记录
General synthesized agentic tasks · 50-task ablation; Pass@2; unit printed as %新增—本次记录
General synthesized agentic tasks · 50-task ablation; Pass@4; unit printed as %新增—本次记录
SWE-bench Verified · Agent mode; task snapshot unspecified本次未见前代记录—
SWE-bench Multilingual · Agent mode; task snapshot unspecified本次未见前代记录—
Terminal-Bench · Terminus 1 framework; task version unspecified本次未见前代记录—

测试记录 · 17