According to PANews, on April 30, Alibaba’s Qwen announced the open-sourcing of Qwen-Scope, an interpretability module trained on Qwen3 and Qwen3.5 series models. The release covers 7 large language models across dense and mixture-of-experts variants, with 14 sets of sparse autoencoder weights.
Disclaimer: The information on this page may come from third parties and does not represent the views or opinions of Gate. The content displayed on this page is for reference only and does not constitute any financial, investment, or legal advice. Gate does not guarantee the accuracy or completeness of the information and shall not be liable for any losses arising from the use of this information. Virtual asset investments carry high risks and are subject to significant price volatility. You may lose all of your invested principal. Please fully understand the relevant risks and make prudent decisions based on your own financial situation and risk tolerance. For details, please refer to
Disclaimer.
Related Articles
DeepSeek Introduces Visual Primitives Method to Enhance Multimodal Reasoning on April 30
According to DeepSeek's technical report, on April 30, the company introduced Visual Primitives, a method that embeds basic visual units such as points and bounding boxes into reasoning chains to address the Reference Gap problem in multimodal tasks. The method reduces image token consumption
GateNews16m ago
NVIDIA Releases Cosmos-Reason2-32B Flagship Model Weights, Expands Context Window to 256K Tokens
According to Beating, NVIDIA has released the weights for Cosmos-Reason2-32B, the flagship version of its physical AI reasoning vision-language model (VLM) designed to help robots and autonomous driving systems understand spatial, temporal, and physical principles. The 32-billion-parameter model,
GateNews18m ago
OpenAI reveals why Codex is not allowed to talk about “goblins”: the nerd persona reward went out of control
OpenAI’s official blog explains that Codex bans “banter goblins” and other creatures because the reward signal in nerd-persona training favored biological metaphors, leading to cross-persona contamination and RLHF misdirection. The incident was revealed by Barron Roth after a system prompt surfaced; OpenAI then used two strategies—short-term hard-coded fixes and long-term reward-signal removal—to warn about the fragility of reward design, and said post-training audits need to be more granular.
ChainNewsAbmedia1h ago
Nvidia B300 AI Servers Hit $1 Million in China Amid Supply Squeeze
According to Reuters, Nvidia's B300 AI servers now sell for approximately 7 million yuan (US$1 million) in China, driven by a crackdown on smuggling networks and continued demand from local tech firms. The price has surged from about 4 million yuan (US$585,000) in late 2025, significantly
GateNews1h ago
LG CNS Operating Profit Rises 19% in Q1 on AI and Cloud Growth
According to Chosun Daily, LG CNS reported on April 30 that first quarter operating profit increased 19.4% year on year to 94.2 billion won (US$64.1 million), driven by AI and cloud demand. Revenue rose 8.6% to 1.3 trillion won (US$894 million).
The company's AI and cloud unit generated 765.4 billi
GateNews2h ago
OpenAI Releases GPT-5.5-Cyber: Battles Anthropic Mythos
OpenAI announced the launch of GPT-5.5-Cyber, designed specifically for cybersecurity, and will deploy it in “the coming days” for key defensive units via a trusted access mechanism in collaboration with the government, not for open sale. In contrast to Anthropic’s Mythos open path, the White House opposes expanding it to 70 organizations. The Preparedness Framework rated it High (not yet critical), while strengthening cybersecurity protections. Going forward, attention will be on the deployment roster, whether CISA could be added, and the differences in security standards between the two sides.
ChainNewsAbmedia3h ago
