| 规格 | 数值 |
|---|---|
| 发布方 | DeepSeek[1][3] |
| 参数量 | 685B(Hugging Face[2] 元数据) |
| 许可证 | MIT |
| 架构 | 与 V3.2-Exp 相同,采用 DeepSeek[1][3] Sparse Attention |
| 工具调用 | V3.2:支持,思考与非思考模式均可;V3.2-Speciale:不支持 |
| 获取方式 | V3.2 提供 App、网页和 API;V3.2-Speciale 仅通过临时端点提供 API,至 2025 年 12 月 15 日 |
| 推荐采样 | temperature 1.0,top_p 0.95 |
来自 Hugging Face 模型卡和 DeepSeek 发布文章;两者均未给出上下文长度或 V3.2 的 API 价格。图表数据为 DeepSeek 发布时的对比表。 · 来源
Estimate from the version's published specs, benchmarks and reviews, 0-5; scores are only comparable within one criteria year. Scored as launched; flagship DeepSeek-V3.2 (1 Dec 2025), Speciale was API-only on a temporary endpoint (no tool calls) at the same price. Reasoning null: AA's index version at launch was retired (AA's later 2026 index shows 21 estimated, 42 on v4.0 per AA's V4 article, neither a launch reading); DeepSeek claimed GPT-5 level and Speciale IMO/IOI gold. Coding null. Context 128K and $0.28 in / $0.42 out (pricing page, Wayback 5 Dec 2025; about $0.32 blended). Speed null. Text only. Reviews: Simon Willison, Composio test (DEV), DEV Community overview, eesel (analysis). api-docs.deepseek.com
Here's what I got from the Speciale model, which thought deeply about the geometry of bicycles and pelicans for a very long time (at least 10 minutes) before spitting out this result
DeepSeek V3.2 is far cheaper but noticeably slower and requires more manual fixes.
While still behind GPT-5 in some multi-step coding tasks, V3.2 demonstrates its strengths in bug fixing and code generation
The official technical report shows that the standard DeepSeek V3.2 model performs on par with GPT-5 on key benchmarks.
| DeepSeek-V3.2 · 2025年12月1日 | 0.10 |
|---|---|
| How DeepSeek V3.2 is Redefining Open-Source AI: GPT-5 and Gemini Challengers · 2025年12月5日 | 0.30 |
| Deepseek 3.2 Speciale vs Claude 4.5 Sonnet for agentic coding · 2025年12月16日 | -0.30 |
| An overview of DeepSeek V3.2: Features, performance, and what it means for AI · 2026年1月6日 | 0.50 |
读者上传的照片,与图钉带来源的媒体分开显示。你可以为吃到或看到的内容评分,在餐厅图钉上还能直接在照片上标记菜品。
第一项始终是图钉的来源。总体可信度是各份资料对上方所用开始和结束时间支持程度的加权平均;资料每比最新的一份旧 180 天,权重减半。
DeepSeek[3]的帖子“DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models”日期为2025年12月1日,称V3.2“now live on App, Web & API”;其更新日志列出“Date: 2025-12-01 DeepSeek-V3.2”,ETV Bharat[6] 12月2日的报道也引用了DeepSeek 12月1日的发布帖。
DeepSeek's[1][3] model card: MIT license, DeepSeek Sparse Attention, a scalable RL framework that makes V3.2 comparable to GPT-5 and V3.2-Speciale surpass it, gold-medal IMO and IOI 2025 performance and the agentic task-synthesis pipeline.
DeepSeek's[1] changelog entry for 2025-12-01: deepseek-chat and deepseek-reasoner upgraded to V3.2 (non-thinking and thinking), and Speciale served from a temporary endpoint with V3.2 pricing, no tool calls, until 15 December 2025 15:59 UTC.
OpenRouter's model record: released 2025-12-01, 163,840-token context, third-party provider prices starting at about $0.21 per million input tokens and $0.31 per million output tokens.
Technical report submitted 2 December 2025 by DeepSeek[1][3]-AI: describes DSA, the RL protocol with scaled post-training compute and the large-scale agentic synthesis pipeline behind V3.2 and Speciale.
有遗漏或错误吗?用你自己的话说明:能佐证此图钉的链接、不同的开始或结束日期及理由,或缺失、有误的信息。AI 会对照此图钉的来源进行核实,搜索更好的来源,并添加任何支持你说法的页面。图钉自身的来源仍然最重要。AI 也会查看图片:显示的是别的东西或显示效果差的图片会被移到后面或替换。
原标题: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale Released
Epoch AI model database (DeepSeek-V3.2); Epoch's confidence in the figures: Likely. Blank tiles are not published.