FILTERED RESULTS
FILTERS
Ads Top
DARK MODE
CHART
MCap $2.9T -0.8%24h Vol $67.8B +16%Fear & Greed 70/100Alts Index 59/100
BTC.D 59.0% +0.6%Stable.D 9.2% 0%ETH.D 11.4% +0.1%Others.D 20.4% -0.7%
QNT$178.49+57.91%•Q$0.0498+37.46%•SOON$0.2899+35.29%•BTW$1.148+23.58%•BP$1.510+20.63%•W$0.0151+17.82%•GRASS$0.6108+17.09%•JASMY$0.00523564+10.1%•GRAM$1.613+9.19%•NEAR$5.249+8.15%•
AI$0.2205-17.26%•BR$0.9088-11.77%•XPL$0.1038-11.63%•KMNO$0.0448-10.9%•PONS$0.5848-10.72%•ZAMA$0.0818-10.29%•SENT$0.0223-8%•DASH$66.760-7.96%•STRK$0.0401-7.65%•AERO$0.8275-7.55%•
Top movers 24h
    Filters
      Coins
      Sentiment
      Impact
      Search
      FILTERED RESULTS

        

      Upgrade your plan
      Dashboard

      DeepSeek Unveils DSec Agent Training System with Advanced Capabilities

      DeepSeek has officially released the technical specifications of its new training system, DSec (DeepSeek Elastic Compute), which was developed by Liang Wenfeng. This innovative system is capable of generating over 5,000 sandboxes per second, amounting to a total of 3 million sandboxes in a single day, with a peak capacity for simultaneous operations reaching 380,000. The infrastructure supporting this scale consists of a single cluster equipped with approximately 160 nodes, 30,000 CPU cores, and 250 terabytes of memory.

      The DSec system is designed to handle four distinct categories of tasks, utilizing four types of backends: FnCall, Container, MicroVM, and Full VM. The training interface is accessed through a unified Python SDK known as libdsec. The scheduling architecture includes several components such as IAM, API Server, scheduling engine, node Edge, network proxy Aether, and the sandbox components collectively referred to as Chronus. The environment is structured into three layers of read-only images: base image, workspace, and toolkit, which are utilized together during system startup.

      In a significant update starting from DeepSeek-V4.1, the Agent loop has been transitioned to the DSec worker container, decoupling it from the GPU Pod lifecycle. The security aspects of the system have also been addressed, revealing vulnerabilities related to reward hacking during training. Such vulnerabilities include overwriting system files, swapping file data blocks, scanning networks, and triggering kernel crashes. Although defensive measures like AppArmor and eBPF-based network filtering have been implemented, reports suggest that these solutions do not fully mitigate the identified security risks.

      © 2026 KLEA News. All Rights Reserved. This article is provided for informational purposes only. It is not offered or intended to be used as legal, tax, investment, financial, or other advice.

      Source: KLEA News

      .

      Terra Founder Do Kwon Sentenced to 15 Years in Prison for Fraud