DeepSeek V4
PulseAugur coverage of DeepSeek V4 — every cluster mentioning DeepSeek V4 across labs, papers, and developer communities, ranked by signal.
- developed by DeepSeek 100%
- subsidiary of DeepSeek 100%
- used by 36Kr 90%
- used by Huawei Ascend 90%
- used by Moore Threads 90%
- developed by DeepSeek V4-Flash 90%
- instance of DeepSeek V4-Flash 90%
- founded Liang Wenfeng 90%
- competes with Moonshot AI 80%
- used by SGLang 80%
- instance of DeepSeek 70%
- used by GitHub 70%
- 2026-05-25 research_milestone DeepSeek V4 completes full adaptation to Huawei Ascend chips, marking a milestone for China's domestic AI stack. 来源
- 2026-05-24 research_milestone DeepSeek V4 is presented as a new AI model challenging OpenAI.
- 2026-05-23 product_launch DeepSeek released its V4 model with a 1 million token context window. 来源
- 2026-05-18 product_launch DeepSeek launched its V4 models, V4-Pro and V4-Flash, on April 24, 2026.
- 2026-05-16 research_milestone DeepSeek V4 achieves a 98% reduction in KV-cache memory usage with its new compressed attention architecture. 来源
- 2026-05-16 product_launch DeepSeek V4, a new open-weight LLM family, was released with significant architectural improvements and cost reductions. 来源
- 2026-05-15 product_launch DeepSeek released its V4 model with MegaMoE optimizations. 来源
- 2026-05-11 product_launch DeepSeek V4, an AI model with a sparse Mixture-of-Experts architecture, was released.
- 2026-05-11 product_launch DeepSeek officially released its new flagship model, DeepSeek-V4. 来源
13 天有情绪数据
-
DeepSeek V4 to launch late April with trillion parameters, Huawei Ascend chip support
DeepSeek founder Liang Wenfeng has revealed that the company's next-generation flagship model, DeepSeek V4, is slated for release in late April. This new model is expected to feature trillion-scale parameters and a mill…
-
Hugging Face Transformers 库新增模型并修复错误
Hugging Face 的 `transformers` 库发布了一系列版本和补丁，引入了新模型并修复了各种错误。值得注意的是，5.9.0 版本添加了 Cohere 的 Command A+ (Cohere2Moe) 和 HRM-Text，同时改进了音频支持和生成能力。早期版本，如 v5.8.0，集成了 DeepSeek-V4、Gemma 4 Assistant、GraniteSpeechPlus、Granite4Vision、EX…
-
Moonshot Kimi K2.5 - 以一半的成本击败 Sonnet 4.5，SOTA开源模型，首个原生图像+视频，100个并行Agent Swarm管理器
Moonshot 发布了 Kimi K2.6，一个更新的开源模型，增强了其在代理编码和多模态理解方面的能力。新版本采用 1T 参数的混合专家架构，拥有 32B 激活参数和 384 个专家，支持 256K 上下文窗口以及原生图像/视频处理。Kimi K2.6 在各种编码和推理基准测试中声称达到了最先进的性能，包括具有数千次工具调用和扩展自主运行的长视距任务。
-
Databricks 使用 Hydra 扩展监控；nOps 在 Lakebase 上重建
Databricks 开发了一个名为 Hydra 的新监控平台，构建在其 Lakehouse 架构之上，以处理其海量运营规模，每日摄取超过 10 万亿个样本并管理 50 亿个活跃时间序列。该平台解决了高基数指标的挑战，并旨在实现更少干预、自我修复的基础设施。同时，nOps 使用 Databricks Lakebase 重建了其云优化平台，整合了其应用程序和分析，以实现更简单、更快的架构。此外，多家公司正在推出旨在简化跨 AWS、GCP…