News Heat Trend
Project Overview
Golem is a peer-to-peer, decentralized computation network that enables a marketplace for computing power. It allows anyone to share and aggregate computing resources, creating a network to share resources. By doing so, Golem hopes to provide software developers with an alternative to traditional centralized cloud service providers, such as Amazon.
B.AI’s “Self-Selected Service Provider” Program Expands Again: Access to Leading Models + Four-Tier Discount Plans Enable Compute Freedom
The B.AI platform’s “Self-Selected Service Provider” model matrix has officially expanded, newly integrating leading large language models including Moonshot (Kimi series) and Z.ai (GLM series). This module offers four discount tiers: 90%, 60%, 40%, and 20% off. Users can generate a personalized “discounted model API key” with a single click, enabling seamless switching between core business operations and routine testing—achieving an optimal balance of high availability and low cost. Moreover, all discounts can be stacked with up to a 1:1 top-up bonus, further lowering the barrier to compute access. Starting today, log in to the B.AI console to customize your专属 model portfolio and enter a new era of AI API calls delivering unmatched value.
rsETH Hack Causes 68,900 ETH Shortfall; DeFi United Raises 13,500 ETH for Industry自救
According to on-chain analyst Ember (@EmberCN), the rsETH incident on April 18 resulted in a funding shortfall of approximately 68,900 ETH (around $160 million): the hacker collateralized rsETH to borrow 99,600 ETH; after Arbitrum recovered 30,700 ETH, the remaining funds were fully converted by the hacker into BTC. The incident has now entered the remediation phase. Aave is coordinating the establishment of a “DeFi United” relief fund, which has so far received cumulative donations totaling 13,500 ETH (approximately $31.45 million). Donors include Lido Finance (2,500 stETH), ether.fi Foundation (5,000 ETH), Aave founder Stani Kulechov (5,000 ETH), Golem Foundation (1,000 ETH), as well as LayerZero and Ink Foundation (amounts undisclosed).
US Department of Commerce Evaluates Kimi K3, Claims U.S. Still Leads, But Report Notes Test Was Not Fully Equivalent
the U.S. Department of Commerce's AI Standards and Innovation Center, in collaboration with the UK AI Safety Institute, tested the cyber attack capabilities of Kimi K3, emphasizing that "the United States still leads."However, the value of the evaluation is debated due to limitations in the testing scope. Due to hosting environment constraints, Kimi K3 only participated in partial testing, with its overall cyber capabilities estimated primarily based on 41 exploit benchmarks. In contrast, other models underwent more comprehensive testing, resulting in a larger margin of error for Kimi K3's results.In the exploit testing, Kimi K3 scored approximately 32%, higher than GLM-5.2's 24%, but lower than the average of approximately 76% for leading U.S. models. In a simulated attack chain test, Kimi K3 completed an average of 17 out of 32 steps in the attack chain and successfully breached the network once in 10 attempts, while U.S. frontier models completed an average of 28.5 steps.The report notes that Kimi K3 already possesses a certain level of autonomous attack capability, and its security guardrails did not prevent the model from developing exploits or executing attacks. However, the report also emphasizes that the testing scope was limited.
OpenAI Model Breaches Test Sandbox and Infiltrates Hugging Face Production Infrastructure to Obtain Benchmark Answers
OpenAI confirmed that the unreleased GPT-5.6 Sol and another unnamed, more powerful pre-release model breached a restricted sandbox environment during ExploitGym benchmark evaluations and infiltrated Hugging Face's production infrastructure to obtain test answers.OpenAI stated that the models leveraged a zero-day vulnerability in an internal software package registry proxy to escalate privileges and move laterally, ultimately connecting to a machine with internet access. The models then identified and chained together vulnerabilities in both the OpenAI research environment and Hugging Face's production infrastructure, directly retrieving test solutions from Hugging Face's production database.Hugging Face disclosed the incident on July 16, stating that the attack was executed end-to-end by an autonomous AI agent system, involving thousands of operations within short-lived sandboxes and accessing internal datasets and service credentials. OpenAI confirmed its models were the subject of the incident five days later.Hugging Face stated that its security team, in order to analyze over 17,000 attack logs, initially attempted to use a commercial US frontier AI interface, but the request was blocked due to safety guardrails. They subsequently switched to using the 753-billion parameter open-weight model GLM 5.2 from Chinese AI startup Z.ai on their own infrastructure to complete the forensic analysis.
Dragonfly Partner: No "Hacker Apocalypse" in DeFi; Annualized Stolen Value in 2026 Estimated at $1.89 Billion
Haseeb posted on X, stating that with models like GLM 5.2, Fable, and GPT 5.6 already launched and actively used by attackers, DeFi has not experienced the anticipated "hacker apocalypse." Chart data shows that based on the current year's data and running rate, the annualized amount stolen from DeFi in 2026 is approximately $1.89 billion. The cumulative stolen amount for the year is around $986 million, lower than the 2025 level and still within the historical range. Haseeb noted that the deeper change now is that while the number of hacker attacks has increased, the scale of individual attacks is declining more rapidly. Attackers are increasingly targeting smaller protocols and abandoned projects, while large protocols have implemented more security enhancements. As a result, overall fund security has not significantly deteriorated.
rsETH Hack Causes 68,900 ETH Shortfall; DeFi United Raises 13,500 ETH for Industry自救
According to on-chain analyst Ember (@EmberCN), the rsETH incident on April 18 resulted in a funding shortfall of approximately 68,900 ETH (around $160 million): the hacker collateralized rsETH to borrow 99,600 ETH; after Arbitrum recovered 30,700 ETH, the remaining funds were fully converted by the hacker into BTC. The incident has now entered the remediation phase. Aave is coordinating the establishment of a “DeFi United” relief fund, which has so far received cumulative donations totaling 13,500 ETH (approximately $31.45 million). Donors include Lido Finance (2,500 stETH), ether.fi Foundation (5,000 ETH), Aave founder Stani Kulechov (5,000 ETH), Golem Foundation (1,000 ETH), as well as LayerZero and Ink Foundation (amounts undisclosed).
Perplexity CEO: GLM 700B Parameter Model Performance Close to Opus Level
Perplexity CEO Aravind Srinivas tweeted that GLM is a severely underrated model, demonstrating performance close to Opus-level at the 700B parameter scale with extremely high operational efficiency. Previously, Moonshot AI just released the open-source model Kimi K3, and industry attention on the GLM series models under Zhipu AI is rising.
Kimi K3 Released, Gap Between Open-Source and Closed-Source Models Narrows to 4 Points
Moonshot AI launches open-source model Kimi K3, scoring 57 points on the Artificial Analysis Intelligence Index, becoming the third highest-scoring model, second only to Anthropic's Claude Opus 5 (61 points), Claude Fable 5 (60 points), and OpenAI's GPT-5.6 Sol (59 points). The index shows that the gap between leading closed-source models and open-weight models has narrowed to 4 points, the smallest gap since the release of GLM-5 in February. This signifies that open-source models are rapidly catching up to closed-source models in performance.
Moonshot AI to Open Source 2.8-Trillion-Parameter Kimi K3 Weights, Chinese Open-Weight Model Token Share Rises to 68%
Chinese AI startup Moonshot AI will release the model weights of its high-performance model, Kimi K3. Developers can download the model, modify it for various purposes, and run it in their own data centers or cloud environments.Kimi K3 boasts 2.8 trillion parameters and a 1 million token context window, enabling it to process large-scale documents and codebases in a single pass. Moonshot AI plans to later publish a technical report detailing the model's architecture, training methodology, and performance evaluation results.Following the release of Kimi K3, Moonshot AI's daily revenue is reported to have increased by at least 6 times. The company is reportedly advancing a new round of fundraising at a $50 billion valuation and is considering a Hong Kong listing as early as this year.According to Bloomberg Intelligence, following the release of Kimi K3 and Z.AI's GLM-5.2, the share of Chinese open-weight models in overall token usage has risen to 68%. Services like AWS Bedrock, Microsoft Azure Foundry, and Google Vertex AI currently do not offer Chinese open-weight models such as Kimi K3 and GLM-5.2.
OpenAI Model Breaches Test Sandbox and Infiltrates Hugging Face Production Infrastructure to Obtain Benchmark Answers
OpenAI confirmed that the unreleased GPT-5.6 Sol and another unnamed, more powerful pre-release model breached a restricted sandbox environment during ExploitGym benchmark evaluations and infiltrated Hugging Face's production infrastructure to obtain test answers.OpenAI stated that the models leveraged a zero-day vulnerability in an internal software package registry proxy to escalate privileges and move laterally, ultimately connecting to a machine with internet access. The models then identified and chained together vulnerabilities in both the OpenAI research environment and Hugging Face's production infrastructure, directly retrieving test solutions from Hugging Face's production database.Hugging Face disclosed the incident on July 16, stating that the attack was executed end-to-end by an autonomous AI agent system, involving thousands of operations within short-lived sandboxes and accessing internal datasets and service credentials. OpenAI confirmed its models were the subject of the incident five days later.Hugging Face stated that its security team, in order to analyze over 17,000 attack logs, initially attempted to use a commercial US frontier AI interface, but the request was blocked due to safety guardrails. They subsequently switched to using the 753-billion parameter open-weight model GLM 5.2 from Chinese AI startup Z.ai on their own infrastructure to complete the forensic analysis.
智谱建成 1GW 国产芯片数据中心，强化训练算力布局
据彭博社报道，智谱 AI 已完成一座仅采用国产芯片的数据中心建设，并已开始部分投运。该数据中心设计功率为 1 吉瓦，将用于支持其 GLM 大模型 的训练与研发。报道称，智谱 AI 目前已建成或运营多个计算集群，单个集群芯片数量均超过 1 万枚。
Goldman Sachs: After Kimi K3 Release, Zhipu and MiniMax Plummet, Era of Compute Monopoly Faces End
According to TechFlow Research, Goldman Sachs' July 18 report pointed out that Moonshot AI released the Kimi K3 model, with 2.8 trillion parameters, surpassing Claude Fable 5 and GPT-5.6 Sol to top the Arena.ai coding leaderboard, with API pricing at $2.3 per million tokens setting a new high for Chinese models. Two days after the release, Zhipu AI fell 28%, MiniMax fell 16%, Nasdaq 100 index futures fell over 1.8%, and the Philadelphia Semiconductor Index cumulatively fell over 18% from highs. Goldman Sachs believes Kimi K3 marks a turning point: a Chinese lab unable to match the largest pre-training compute capacity in the West rapidly narrowed the gap with top US models through architectural innovation and reinforcement learning, proving that "scaling" is no longer the only winning path. Goldman Sachs warns that the "compute expansion era" may be ending, and the AI infrastructure investment logic built around "the more compute, the better" needs to be rewritten. Goldman Sachs maintains a Buy rating on MiniMax and Neutral on Zhipu AI. Future focus should be on the intensive launch of 2-5 trillion parameter models such as Zhipu GLM, Alibaba Qwen, and MiniMax M3 Pro.
Related news
Perplexity CEO: GLM 700B Parameter Model Performance Close to Opus Level
Perplexity CEO Aravind Srinivas tweeted that GLM is a severely underrated model, demonstrating performance close to Opus-level at the 700B parameter scale with extremely high operational efficiency. Previously, Moonshot AI just released the open-source model Kimi K3, and industry attention on the GLM series models under Zhipu AI is rising.
Kimi K3 Released, Gap Between Open-Source and Closed-Source Models Narrows to 4 Points
Moonshot AI launches open-source model Kimi K3, scoring 57 points on the Artificial Analysis Intelligence Index, becoming the third highest-scoring model, second only to Anthropic's Claude Opus 5 (61 points), Claude Fable 5 (60 points), and OpenAI's GPT-5.6 Sol (59 points). The index shows that the gap between leading closed-source models and open-weight models has narrowed to 4 points, the smallest gap since the release of GLM-5 in February. This signifies that open-source models are rapidly catching up to closed-source models in performance.
Moonshot AI to Open Source 2.8-Trillion-Parameter Kimi K3 Weights, Chinese Open-Weight Model Token Share Rises to 68%
Chinese AI startup Moonshot AI will release the model weights of its high-performance model, Kimi K3. Developers can download the model, modify it for various purposes, and run it in their own data centers or cloud environments.Kimi K3 boasts 2.8 trillion parameters and a 1 million token context window, enabling it to process large-scale documents and codebases in a single pass. Moonshot AI plans to later publish a technical report detailing the model's architecture, training methodology, and performance evaluation results.Following the release of Kimi K3, Moonshot AI's daily revenue is reported to have increased by at least 6 times. The company is reportedly advancing a new round of fundraising at a $50 billion valuation and is considering a Hong Kong listing as early as this year.According to Bloomberg Intelligence, following the release of Kimi K3 and Z.AI's GLM-5.2, the share of Chinese open-weight models in overall token usage has risen to 68%. Services like AWS Bedrock, Microsoft Azure Foundry, and Google Vertex AI currently do not offer Chinese open-weight models such as Kimi K3 and GLM-5.2.
US Department of Commerce Evaluates Kimi K3, Claims U.S. Still Leads, But Report Notes Test Was Not Fully Equivalent
the U.S. Department of Commerce's AI Standards and Innovation Center, in collaboration with the UK AI Safety Institute, tested the cyber attack capabilities of Kimi K3, emphasizing that "the United States still leads."However, the value of the evaluation is debated due to limitations in the testing scope. Due to hosting environment constraints, Kimi K3 only participated in partial testing, with its overall cyber capabilities estimated primarily based on 41 exploit benchmarks. In contrast, other models underwent more comprehensive testing, resulting in a larger margin of error for Kimi K3's results.In the exploit testing, Kimi K3 scored approximately 32%, higher than GLM-5.2's 24%, but lower than the average of approximately 76% for leading U.S. models. In a simulated attack chain test, Kimi K3 completed an average of 17 out of 32 steps in the attack chain and successfully breached the network once in 10 attempts, while U.S. frontier models completed an average of 28.5 steps.The report notes that Kimi K3 already possesses a certain level of autonomous attack capability, and its security guardrails did not prevent the model from developing exploits or executing attacks. However, the report also emphasizes that the testing scope was limited.
OpenAI Model Breaches Test Sandbox and Infiltrates Hugging Face Production Infrastructure to Obtain Benchmark Answers
OpenAI confirmed that the unreleased GPT-5.6 Sol and another unnamed, more powerful pre-release model breached a restricted sandbox environment during ExploitGym benchmark evaluations and infiltrated Hugging Face's production infrastructure to obtain test answers.OpenAI stated that the models leveraged a zero-day vulnerability in an internal software package registry proxy to escalate privileges and move laterally, ultimately connecting to a machine with internet access. The models then identified and chained together vulnerabilities in both the OpenAI research environment and Hugging Face's production infrastructure, directly retrieving test solutions from Hugging Face's production database.Hugging Face disclosed the incident on July 16, stating that the attack was executed end-to-end by an autonomous AI agent system, involving thousands of operations within short-lived sandboxes and accessing internal datasets and service credentials. OpenAI confirmed its models were the subject of the incident five days later.Hugging Face stated that its security team, in order to analyze over 17,000 attack logs, initially attempted to use a commercial US frontier AI interface, but the request was blocked due to safety guardrails. They subsequently switched to using the 753-billion parameter open-weight model GLM 5.2 from Chinese AI startup Z.ai on their own infrastructure to complete the forensic analysis.
智谱建成 1GW 国产芯片数据中心，强化训练算力布局
据彭博社报道，智谱 AI 已完成一座仅采用国产芯片的数据中心建设，并已开始部分投运。该数据中心设计功率为 1 吉瓦，将用于支持其 GLM 大模型 的训练与研发。报道称，智谱 AI 目前已建成或运营多个计算集群，单个集群芯片数量均超过 1 万枚。