Model dossier模型档案
DeepSeek V3.2
DeepSeek V3.2 is a mid-tier model from DeepSeek (Intelligence Score 69.9/100). Strongest in knowledge (89.5) and agent (86.2); best fit: knowledge QA and long-form writing. It trails in preference (31.5). Budget-priced: $0.269 in / $0.400 out per 1M tokens. Open weights — self-hostable. Reachable via CN-direct endpoints.
DeepSeek V3.2 是 DeepSeek 的中端梯队模型(智能评分 69.9/100)。 最强项是知识(89.5分),其次是智能体(86.2分);适合知识问答与长文写作。 短板是偏好(31.5分)。 定价属低价档:每百万 token 输入 $0.269 / 输出 $0.400。 开放权重,可自托管。 支持国内直连。
open weights开放权重 ⌂ CN direct国内直连
69.9
Intelligence Score · benchmark coverage 智能评分 · benchmark 覆盖 100%
◈
Benchmark scoresBenchmark 成绩
11 entries 条| Benchmark基准 | Score成绩 | Index指数 | Dated日期 | Measured by测评方 | |
|---|---|---|---|---|---|
| Humanity's Last ExamReasoning | 25.1 | 44 | — | 3rd-party第三方 | source ↗×3 |
| GPQA DiamondReasoning | 82.9 | 87 | — | 3rd-party第三方 | source ↗×4⚠±5.4 |
| ↳ ⚡high | 82.4 | — | 2025-12 | vendor-reported厂商自报 | source ↗ |
| MMLU-ProKnowledge | 85 | 89 | — | 3rd-party第三方 | source ↗×3 |
| SWE-bench VerifiedCoding | 73.1⚡high | 76 | 2026-08 | 3rd-party第三方 | source ↗×7 |
| AIME 2025Math | 96⚡high | 96 | — | 3rd-party第三方 | source ↗×4⚠±13.6 |
| LiveCodeBenchCoding | 89.6 | 96 | 2026-07 | 3rd-party第三方 | source ↗×4⚠±6.3 |
| FrontierMathMath | 2⚡high | 2 | — | 3rd-party第三方 | source ↗×2⚠±20.1 |
| Terminal-BenchCoding | 46.4 | 50 | — | official官方榜 | source ↗×2 |
| τ²-benchAgent | 91⚡high | 92 | 2025-12-01 | 3rd-party第三方 | source ↗×3⚠±5.2 |
| LMArena (Chatbot Arena)Preference | 1425 | 32 | 2026-08-12 | official官方榜 | source ↗×4⚠±43 |
| ↳ ⚡off | 1425 | — | — | 3rd-party第三方 | source ↗ |
| WebArenaAgent | 74.3 | 81 | 2026-02 | 3rd-party第三方 | source ↗×2 |
≡
Specs & pricing规格与价格
| Provider厂商 | DeepSeek |
|---|---|
| Released发布 | 2025-12-01 |
| Context window上下文 | 164K tokens |
| Max output最大输出 | 66K tokens |
| Modality模态 | text->text |
| Reasoning tiers推理档位 | ⚡none ⚡high default默认 none · thinking on / off |
| API input price输入价格 | $0.269 / 1M tokens |
| API output price输出价格 | $0.400 / 1M tokens |
⛓
Tools using this model使用它的工具
How we score →评分方法 → · Catalog & pricing via the official OpenRouter API.模型目录与价格来自 OpenRouter 官方 API。