可进入对比
档案摘要 / Decision snapshot
先判断能不能用于选型。
有真实任务证据
来自模型卡
边界优先
30 秒结论
可进入选型对比,但仍以案例证据为准。
来自公开模型资料库,并已补厂商/官方/模型家族证据;当前站内暂无可核验 A 类案例。
这页有来源、有风险说明、有公开可核验 A 类案例。
数据可信度 / 证据强度
强
A 类案例
5
AA 评分
12.51
官方来源
3 个补充入口
案例来源
5 条可核验
来自公开资料与模型数据库 + Artificial Analysis + 3 个厂商/官方/模型家族证据入口;AA 不是唯一事实来源。
Evidence Distribution
案例证据分布
DeepSeek-V2.5 (Dec) 当前关联 5 条 A 类案例;这里按任务、来源和复核时间观察证据结构。
Evidence Snapshot
证据快照覆盖
DeepSeek-V2.5 (Dec) 的 5 条 A 类案例中,已有 5 条完成本地证据快照;0 条仍在快照队列。
可用 manifest 复核
等待落盘
需补抓
需人工复核
证据完整度
5 / 5
Feishu Bitable
3 个入口
5 条
已绑定
已标注
能力边界
用标签表达信号,不用标签替代证据。
通用/待核验
暂无数据
待核验
开放/混合信号
中
| 厂商 | DeepSeek |
|---|---|
| 厂商页 | DeepSeek |
| 发布时间 | 2024-12-10 |
| 模型 ID | deepseek-v2-5-dec |
| 上下文 | 暂无数据 |
| 输出 | 暂无数据 |
| 模态 | 文本 / 多模态待核验 |
| 推理 | 否或官方未披露 |
| 价格 | 官方未披露 / 暂无数据 |
| 可用平台 | DeepSeek |
| AA 评分 | 12.51 |
| 基础模型发布时间 | Artificial Analysis 未披露 |
| 资料来源 | 公开资料、厂商信息和案例库 |
| A 类案例 | 5 |
谱系位置
它在厂商路线中的位置
此页将 DeepSeek-V2.5 (Dec) 放入 DeepSeek 的当前 Atlas 路线中。若同厂前后代资料不足,先保留为路线追踪入口,不脑补谱系关系。
2024-12-10
DeepSeek-V2.5 (Dec) 资料状态
来自公开资料与模型数据库 + Artificial Analysis + 3 个厂商/官方/模型家族证据入口;AA 不是唯一事实来源。
公开档案
模型状态
DeepSeek-V2.5 (Dec) 当前标记为已有真实案例。
案例库
公开案例复核
已有 5 条可核验 A 类案例,达到完整补齐线
案例时间线
最近进入档案的 A 类案例
用采集时间呈现证据进入 Atlas 的顺序,方便复核来源新鲜度。
2026-06-27T02:50:00Z
KTransformers 用户在消费级硬件上本地部署 DeepSeek-V2.5-1210 236B 实现高效推理
liuqjox (via kvcache-ai/ktransformers) / local_inference_deployment
2026-06-26T18:22:09.049999+00:00
CleanS2S: DeepSeek-V2.5 as LLM Backend for Streaming Speech-to-Speech Agent
OpenDILab / speech_to_speech_interaction
2026-06-26T18:22:09.049996+00:00
Raycast G4F Extension: Free DeepSeek-V2.5 Access via DeepInfra in Raycast
XInTheDark / llm_access_and_integration
2026-06-26T18:22:09.049989+00:00
Mem1: DeepSeek-V2.5 as LLM Backend for Emotion Companion Memory Extraction
1596941391qq / conversational_memory_extraction
2026-06-26T18:22:09.049962+00:00
SeekSpider: AI-Powered Australian IT Job Scraping with DeepSeek-V2.5
qinscode / data_extraction_and_analysis
适合场景
- - 作为完整模型索引和厂商路线追踪入口
- - 通用模型对比
不适合
- - 缺少真实案例时不要包装成推荐
- - 缺字段时不要脑补价格、性能或上下文
真实案例
可核验 A 类案例精选
当前展示 5 条代表案例;排序优先看模型卡精选、证据可信度、快照状态和公开仓库/技术任务信号。案例总数仍为 5。
Representative Evidence
代表案例排序
当前第一代表案例是“KTransformers 用户在消费级硬件上本地部署 DeepSeek-V2.5-1210 236B 实现高效推理”。它会优先出现在本页,是因为该案例同时进入模型卡精选、具备 100/100 的证据可信度,并带有 已快照证据。
模型卡精选优先
A+ 完整链路
archive/evidence/ktransformers-deepseek-v25-1210-local-inference/manifest.json
Coding / repo / agent
liuqjox (via kvcache-ai/ktransforme… 使用 DeepSeek-V2.5 (Dec) 处理真实任务执行
liuqjox (via kvcache-ai/ktransformers) · DeepSeek-V2.5 (Dec)
A+ 完整链路 · 代码仓库证据
liuqjox (via kvcache-ai/ktransformers) 公开的真实任务执行案例,来源为 公开代码库,复核于 2026-06-27T02:50:00Z。 Model Atlas 将它标记为 A 类证据,因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。
任务:用户 liuqjox 使用 KTransformers 框架在 Windows/WSL 环境下本地部署量化版 DeepSeek-V2.5-1210 (Q2_K_L 量化, 236B 参数, 80.5GB)。硬件配置为 RTX 3090 + 96GB DDR4 3200 + i7-10700K, 环境为 Win10/WSL/Ubuntu/CUDA 12.4/PyTorch 2.6。目标是在不到一万元的消费级硬件上运行 236B 参数大模型进行本地推理。
公开产物:成功实现本地推理,速度稳定在 4-6 tokens/s(初始较慢约 1 tps)。同时验证了更大量化的 DeepSeek-R1 671B (unsloth 2.51 bit, 212GB) 也可运行但速度极慢(0.05 tps)。用户指出 x99 主板 + 256G 内存方案成本不到一万元即可运行 671B 模型。
模型作用:DeepSeek-V2.5-1210 提供了 236B 参数的开放权重模型,其 MoE 架构配合 KTransformers 的 CPU/GPU 异构推理技术,使得在单卡 RTX 3090 + 96GB 内存的消费级硬件上即可实现可用的本地推理(4-6 tps),证明了该模型在资源受限场景下的实用价值。
A 类理由:有具体使用者、具体任务、公开原始证据和可访问产物,可公开核验。
风险边界:Issue 状态为 open 但用户已确认部署成功;量化版本 Q2_K_L 可能有精度损失;速度数据基于特定硬件配置,不同配置可能差异较大。
原始记录:KTransformers 用户在消费级硬件上本地部署 DeepSeek-V2.5-1210 236B 实现高效推理
OpenDILab 使用 DeepSeek-V2.5 (Dec) 处理文档理解和结构化处理
OpenDILab · DeepSeek-V2.5 (Dec)
A+ 完整链路 · 代码仓库证据
OpenDILab 公开的文档理解与处理案例,来源为 公开代码库,复核于 2026-06-26T18:22:09.049999+00:00。 Model Atlas 将它标记为 A 类证据,因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。
任务:这是一个围绕文档理解和结构化处理的真实任务,公开材料可以回溯到具体使用者和具体产物。 原始资料写作:OpenDILab's CleanS2S project is a high-quality streaming speech-to-speech interactive agent implemented in a single file. DeepSeek-V2.5 is specifically recommended and documented as the default local LLM option for the …
公开产物:公开材料提供公开代码、README 或项目配置,可用于核验任务结果、项目形态和模型绑定关系。 原始资料写作:A fully functional speech-to-speech agent (536 stars) that processes real-time voice input, routes through an LLM for conversation, and generates voice output with prosody transfer. The project supports both API-based a…
模型作用:DeepSeek-V2.5 (Dec) 在该案例中承担文档理解和结构化处理相关的生成、分析、编排或实现角色。 原始资料写作:DeepSeek-V2.5 is the specifically recommended and documented local LLM option for the speech agent's conversational intelligence. The README provides explicit instructions for running V2.5 locally as the LLM backbone, i…
A 类理由:有具体使用者、具体任务、公开原始证据和可访问产物,可公开核验。
风险边界:当前判断基于公开材料;若产物下线、仓库变更或模型参与比例仅来自作者自述,需要在引用前重新复核。 原始资料写作:CleanS2S supports multiple LLM backends; V2.5 is the recommended option, not the only one. The project is from OpenDILab (536 stars), a credible AI research organization. V2.5's role is as the conversational reasoning e…
原始记录:CleanS2S: DeepSeek-V2.5 as LLM Backend for Streaming Speech-to-Speech Agent
XInTheDark 使用 DeepSeek-V2.5 (Dec) 处理真实任务执行
XInTheDark · DeepSeek-V2.5 (Dec)
A+ 完整链路 · 代码仓库证据
XInTheDark 公开的真实任务执行案例,来源为 公开代码库,复核于 2026-06-26T18:22:09.049996+00:00。 Model Atlas 将它标记为 A 类证据,因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。
任务:这是一个围绕真实任务执行的真实任务,公开材料可以回溯到具体使用者和具体产物。 原始资料写作:A Raycast extension that provides free access to multiple LLMs including DeepSeek-V2.5 via DeepInfra. V2.5 is listed as an active, working provider with medium speed and 7.5/10 quality rating. The extension supports cha…
公开产物:公开材料提供公开代码、README 或项目配置,可用于核验任务结果、项目形态和模型绑定关系。 原始资料写作:A Raycast extension (1073 stars) that routes requests to DeepInfra's free tier for DeepSeek-V2.5 access. The model is marked as active (green badge) with medium response speed. Users can access V2.5 for free through the…
模型作用:DeepSeek-V2.5 (Dec) 在该案例中承担真实任务执行相关的生成、分析、编排或实现角色。 原始资料写作:DeepSeek-V2.5 is one of the primary models offered through the extension, listed with active status on DeepInfra. The extension enables zero-cost access to V2.5's general conversation and coding capabilities, making it …
A 类理由:有具体使用者、具体任务、公开原始证据和可访问产物,可公开核验。
风险边界:当前判断基于公开材料;若产物下线、仓库变更或模型参与比例仅来自作者自述,需要在引用前重新复核。 原始资料写作:This is a multi-model access tool, not a V2.5-specific application. V2.5 is accessed through DeepInfra's infrastructure, not DeepSeek's own API. The extension also supports many other models. V2.5's active status was ve…
原始记录:Raycast G4F Extension: Free DeepSeek-V2.5 Access via DeepInfra in Raycast
1596941391qq 使用 DeepSeek-V2.5 (Dec) 处理真实任务执行
1596941391qq · DeepSeek-V2.5 (Dec)
A+ 完整链路 · 代码仓库证据
1596941391qq 公开的真实任务执行案例,来源为 公开代码库,复核于 2026-06-26T18:22:09.049989+00:00。 Model Atlas 将它标记为 A 类证据,因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。
任务:这是一个围绕真实任务执行的真实任务,公开材料可以回溯到具体使用者和具体产物。 原始资料写作:Modified Mem0 framework for emotion companion AI, using DeepSeek-V2.5 as the core LLM to extract and categorize user memories from multi-turn conversations. The system separates temporal memories (e.g., 'went skiing on …
公开产物:公开材料提供公开代码、README 或项目配置,可用于核验任务结果、项目形态和模型绑定关系。 原始资料写作:A working memory extraction pipeline that processes batches of conversation records (recommended 10 at a time), uses V2.5 to extract both user and AI memories, and stores them in a vector database with temporal/constant…
模型作用:DeepSeek-V2.5 (Dec) 在该案例中承担真实任务执行相关的生成、分析、编排或实现角色。 原始资料写作:DeepSeek-V2.5 is configured as the default LLM (model: deepseek-ai/DeepSeek-V2.5 via SiliconFlow) for the memory extraction pipeline. It processes raw conversation transcripts and extracts structured memory entities. Te…
A 类理由:有具体使用者、具体任务、公开原始证据和可访问产物,可公开核验。
风险边界:当前判断基于公开材料;若产物下线、仓库变更或模型参与比例仅来自作者自述,需要在引用前重新复核。 原始资料写作:Repo has 0 GitHub stars. This is a fork/modification of Mem0 (mem0ai/mem0). The developer's customizations focus on Chinese LLM compatibility and emotion companion use case. V2.5 is used via SiliconFlow API proxy.
原始记录:Mem1: DeepSeek-V2.5 as LLM Backend for Emotion Companion Memory Extraction
qinscode 使用 DeepSeek-V2.5 (Dec) 处理研究分析和报告生成
qinscode · DeepSeek-V2.5 (Dec)
A+ 完整链路 · 代码仓库证据
qinscode 公开的研究与报告生成案例,来源为 公开代码库,复核于 2026-06-26T18:22:09.049962+00:00。 Model Atlas 将它标记为 A 类证据,因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。
任务:这是一个围绕研究分析和报告生成的真实任务,公开材料可以回溯到具体使用者和具体产物。 原始资料写作:Automated scraping of IT job listings from Seek.com.au across multiple Australian regions (Perth, Sydney, Melbourne, etc.), with DeepSeek-V2.5 used as the AI backend for post-processing: extracting tech stacks from job …
公开产物:公开材料提供公开代码、README 或项目配置,可用于核验任务结果、项目形态和模型绑定关系。 原始资料写作:A production job scraping pipeline that stores structured job data (title, company, salary, tech stack, region) in PostgreSQL, with AI-extracted fields powered by DeepSeek-V2.5 via SiliconFlow API. The system runs on a …
模型作用:DeepSeek-V2.5 (Dec) 在该案例中承担研究分析和报告生成相关的生成、分析、编排或实现角色。 原始资料写作:DeepSeek-V2.5 is configured as the default AI_MODEL (deepseek-ai/DeepSeek-V2.5 via SiliconFlow) for two core AI tasks: (1) tech_stack_analyzer.py extracts technology keywords from raw HTML job descriptions, and (2) sala…
A 类理由:有具体使用者、具体任务、公开原始证据和可访问产物,可公开核验。
风险边界:当前判断基于公开材料;若产物下线、仓库变更或模型参与比例仅来自作者自述,需要在引用前重新复核。 原始资料写作:The repo has 41 stars and appears to be a personal project. V2.5 is accessed via SiliconFlow proxy, not directly via DeepSeek API. The .env.example explicitly sets AI_MODEL=deepseek-ai/DeepSeek-V2.5 as the default.
原始记录:SeekSpider: AI-Powered Australian IT Job Scraping with DeepSeek-V2.5
数据缺口
只基于现有字段判断缺口;缺失项不会被猜测填充。
- - 补充方向:核验 价格 的官方/API/案例来源。
- - 补充方向:核验 上下文 的官方/API/案例来源。
- - 补充方向:核验 输出 的官方/API/案例来源。
下一步验证建议
- - 继续保留 A 类案例的原始证据、产物页和版本快照。
- - 补官方发布/API/System Card 来源;若官方未披露,继续标注“官方未披露”。
- - 把 benchmark、教程、发布文保留为背景资料,不提升为真实案例。