Author: RadarAI Editorial
Editor: RadarAI Editorial
Last updated: 2026-04-25
Review status: Editorial review pending
Brief
速报
官方
AI动态
开源
The DeepSeek V4 series has officially launched—featuring a 1.6-trillion-parameter Pro version and a 284-billion-parameter Flash version—delivering performance on par with top-tier closed-source models. Notably, it is the first major open model natively optimized for Huawei's Ascend chips, marking a pivotal milestone in China's AI ecosystem's shift away from NVIDIA dependence [11]. Concurrently, the Agent engineering paradigm is accelerating across domains: from intelligent cockpits (by Baizhong, Tencent, and ByteDance) to evaluation frameworks (e.g., Peking University's One-Eval), the 'Model + Harness' approach is supplanting pure model iteration as the core pathway for realizing technical value [12][13][8].
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
## 🔍 Key Insights
**DeepSeek V4** has officially launched—advancing along a dual-track strategy with its **1.6-trillion-parameter Pro version** and **284-billion-parameter Flash version**. Benchmarks confirm performance rivaling leading closed-source models, and it marks the **first major release natively optimized for Huawei's Ascend chips**, establishing a critical milestone in China's AI 'de-NVIDIA-ization' journey [11]. Meanwhile, the **Agent engineering paradigm** is rapidly gaining traction—from intelligent cockpits (Baizhong, Tencent, ByteDance) to evaluation frameworks (Peking University's One-Eval)—as the 'Model + Harness' architecture replaces standalone model iteration as the central path to delivering tangible technical value [12][13][8].
## 🚀 Key Updates
- **DeepSeek V4 debuts on Huawei chips—breaking NVIDIA's dominance in domestic AI** [11]: Both the 1.6T Pro and 284B Flash versions are open-sourced and fully compatible with Ascend 910B, slashing inference costs by over 60%.
- **Peking University open-sources One-Eval: the first automated LLM evaluation framework built on Agents** [8]: Enables natural-language-initiated evaluations, automatic planning of execution paths, and end-to-end traceability—already validated on DeepSeek V4.
- **Baizhong upgrades SuperMate: Intelligent cockpits evolve into on-device Agents** [1]: Achieves full-scenario service orchestration and cross-chip compatibility—shifting user experience from 'command-driven' to 'context-aware proactivity'.
- **Tencent launches seven cockpit agents—redefining automotive time value** [2]: Uses AI to orchestrate information flows, service flows, and payment flows—transforming vehicles into upgradable, revenue-generating mobile intelligent terminals.
- **ByteDance's Doubao Auto 2.0 hits the road: On-device 30B model + cloud-based multi-Agent collaboration** [13]: Supports full-duplex voice interaction and whole-vehicle coordination—with mass production scheduled this year.
- **Granola surpasses $1B valuation: Minimalist design + privacy-first ethos fuel viral growth in AI note-taking** [3]: Zero marketing spend; built entirely on word-of-mouth and end-to-end encryption to establish high-trust barriers.
- **Agent Harness Engineering emerges as a new engineering consensus** [12]: Google Cloud formalizes the formula 'Agent = Model + Harness', emphasizing context engineering, tool integration, and observability as essential 'control-layer' capabilities.
- **Yue Xing Tong's e-paper device secures joint investment from Boyu, Matrix, and Shunwei: A 70g magnetic-mount device pioneers a new AI attention interface** [18]: Monthly sales surged over 10× in six months—defining the 'AI second screen on your phone's back'.
## 🔗 Sources
[1] Baizhong Intelligence SuperMate Upgrade: Intelligent Cockpits Are Evolving from Voice Assistants to On-Device Agents — https://www.bestblogs.dev/article/f13a3132
[2] Tencent Launches Seven Cockpit Agents: The Automobile's Purpose Shifts from 'Spatial Mobility' to 'Time Value' — https://www.bestblogs.dev/article/a9fbbe1a
[3] #507. How Granola's Minimalist Design Earned a $1B Valuation — https://www.bestblogs.dev/podcast/192b198
[8] Within 10 Hours of DeepSeek-V4's Release, Peking University's Open-Source Project Delivers Automated Evaluation! — https://www.bestblogs.dev/article/cb2410b4
[11] Confirmed: DeepSeek V4 Debuts on Huawei Chips! Domestic AI Begins Breaking NVIDIA's 'Monopoly' — https://www.bestblogs.dev/article/c263bec1
[12] Agent Harness Engineering:
DeepSeek V4 has officially launched—advancing along a dual-track strategy with its 1.6-trillion-parameter Pro version and 284-billion-parameter Flash version. Benchmarks confirm performance rivaling leading closed-source models, and it marks the first major release natively optimized for Huawei's Ascend chips, establishing a critical milestone in China's AI 'de-NVIDIA-ization' journey [11]. Meanwhile, the Agent engineering paradigm is rapidly gaining traction—from intelligent cockpits (Baizhong, Tencent, ByteDance) to evaluation frameworks (Peking University's One-Eval)—as the 'Model + Harness' architecture replaces standalone model iteration as the central path to delivering tangible technical value [12][13][8].
🚀 Key Updates
- DeepSeek V4 debuts on Huawei chips—breaking NVIDIA's dominance in domestic AI [11]: Both the 1.6T Pro and 284B Flash versions are open-sourced and fully compatible with Ascend 910B, slashing inference costs by over 60%.
- Peking University open-sources One-Eval: the first automated LLM evaluation framework built on Agents [8]: Enables natural-language-initiated evaluations, automatic planning of execution paths, and end-to-end traceability—already validated on DeepSeek V4.
- Baizhong upgrades SuperMate: Intelligent cockpits evolve into on-device Agents [1]: Achieves full-scenario service orchestration and cross-chip compatibility—shifting user experience from 'command-driven' to 'context-aware proactivity'.
- Tencent launches seven cockpit agents—redefining automotive time value [2]: Uses AI to orchestrate information flows, service flows, and payment flows—transforming vehicles into upgradable, revenue-generating mobile intelligent terminals.
- ByteDance's Doubao Auto 2.0 hits the road: On-device 30B model + cloud-based multi-Agent collaboration [13]: Supports full-duplex voice interaction and whole-vehicle coordination—with mass production scheduled this year.
- Granola surpasses $1B valuation: Minimalist design + privacy-first ethos fuel viral growth in AI note-taking [3]: Zero marketing spend; built entirely on word-of-mouth and end-to-end encryption to establish high-trust barriers.
- Agent Harness Engineering emerges as a new engineering consensus [12]: Google Cloud formalizes the formula 'Agent = Model + Harness', emphasizing context engineering, tool integration, and observability as essential 'control-layer' capabilities.
- Yue Xing Tong's e-paper device secures joint investment from Boyu, Matrix, and Shunwei: A 70g magnetic-mount device pioneers a new AI attention interface [18]: Monthly sales surged over 10× in six months—defining the 'AI second screen on your phone's back'.
🔗 Sources
[1] Baizhong Intelligence SuperMate Upgrade: Intelligent Cockpits Are Evolving from Voice Assistants to On-Device Agents — https://www.bestblogs.dev/article/f13a3132
[2] Tencent Launches Seven Cockpit Agents: The Automobile's Purpose Shifts from 'Spatial Mobility' to 'Time Value' — https://www.bestblogs.dev/article/a9fbbe1a
[3] #507. How Granola's Minimalist Design Earned a $1B Valuation — https://www.bestblogs.dev/podcast/192b198
[8] Within 10 Hours of DeepSeek-V4's Release, Peking University's Open-Source Project Delivers Automated Evaluation! — https://www.bestblogs.dev/article/cb2410b4
[11] Confirmed: DeepSeek V4 Debuts on Huawei Chips! Domestic AI Begins Breaking NVIDIA's 'Monopoly' — https://www.bestblogs.dev/article/c263bec1
[12] Agent Harness Engineering:
← Back to Updates