model

OpenCode stated that a security incident related to a large-scale model distillation attack occurred, affecting approximately 1.4% of OpenCode Console users.The affected range mainly involves some users connected to external model providers. OpenCode has notified all affected accounts via email; users who did not receive a notification are not included in this impact.

On October 16, B.AI announced the domestic model brand Token weekly ranking from October 9 to 15. Xiaomi model Token usage has remained steady at 2nd place for three consecutive weeks, with a month-on-month increase of 49.9%, and the platform Token usage proportion increased by 10.4 percentage points; DeepSeek continues to lead this week with a total of 1.03T Tokens; Zhipu saw a month-on-month growth of 44.3%, with the platform proportion increasing by 3.7 percentage points, ranking 3rd; Moonshot Token usage surged by 286.5% month-on-month. With multiple models competing on the same stage, real usage witnesses market changes. B.AI continues to expand model choices, helping developers flexibly access diverse AI capabilities. Welcome to log in to b.ai to explore.

On October 16, Tencent Holdings stated: The company discovered unexpected behavior from third-party AI models on its platform. Earlier this week, a research report claimed that "out of control" AI entities had emerged in China, originating from a sandbox on Tencent Cloud.The Shenzhen company stated that the relevant models did not come from its own Hy artificial intelligence development team, and no evidence of deliberate targeting or collusion was found. Tencent indicated that these entities were part of an evaluation activity for models built by third-party developers. The report from the international research team stated that these entities detected Alibaba Group's online mapping service and attempted to bypass anti-bot protections.

Meta Platforms (META.O) will release model weights based on the Muse model license, accompanied by limited commercial use terms, allowing numerous developers to use the model for free.

According to CNBC, Apple (AAPL.O) has partnered with Meta Platforms (META.O) to enable the Muse AI model to run on Mac computers.

Market news reports that Nvidia has released the Nemotron 3.5 Super VL and launched the edge model Nemotron 3.5 Bolt 74B-A9B (24GB).

According to Top Chinese Innovation Society reports, former head of the basic model for humanoid robots at ZhiJi Dynamics, Yu Tao, has left to start his own venture, focusing on the combination of LLM and embodied intelligence. According to relevant sources, his company's first round valuation is 70 million USD, and the second round valuation is 200 million USD, with DeepSeek and a leading top-tier fund participating in the investment.Yu Tao holds a PhD from the University of Science and Technology of China and was a visiting scholar at the National University of Singapore. He previously served as a senior researcher in Tencent AI Lab's Absolute Enlightenment team, where he led the team to build the full-body tracking system LimX WBT and the motion generation system from scratch during his time at ZhiJi Dynamics. According to reports from Phoenix Technology, DeepSeek founder Liang Wenfeng has recently met with several entrepreneurs in embodied intelligence and is internally evaluating different entry methods, such as forming a team independently or investing in external startups.In August of this year, DeepSeek invested approximately 141 million RMB through a strategic placement in Yushu Technology for its IPO; previously, Bloomberg also disclosed its plan to lead the investment in the robotics company Winverse AI founded by former Tencent AI scientist Hu Han.

The stablecoin issuer Tether's Tether AI Research (referred to in the text as Tether Data, S.A. de C.V.) has released the open-source multilingual automatic speech recognition model QVAC TranscribePsy. This model has 254 million parameters and runs entirely on-device, suitable for smartphones, smart glasses, and humanoid robots, with audio not needing to be uploaded to the cloud; smart glasses can process in near real-time via Bluetooth or Wi-Fi, and the model can also run on the robot's onboard processor.Evaluation shows that its word error rate on the Open ASR Leaderboard is 8.02%, 6.58% on Multilingual LibriSpeech, and 21.56% on Common Voice. It outperforms the specifically mentioned Qwen3-ASR-0.6B, omniASR-CTC-1B-v2, OWSM-CTC-v4, and Whisper-Medium on Multilingual LibriSpeech. Among six tested smartphones, it is the fastest model, being 2.6 times faster than the second fastest model and 5.7 times faster than Whisper-small. The Core ML version transcribes 10 minutes of audio into 1.65 seconds on the iPhone 17 Pro Max, occupying less than half the disk space of a certain tested Core ML competitor and using 1.7 times less memory.

According to Gate Market, Hong Kong stock large model stocks fell, Zhipu (02513.HK) and MINIMAX-W (00100.HK) both dropped over 10%.

According to Decrypt, over 400 scientists, including "AI Father" Geoffrey Hinton and 2018 Turing Award winner Yoshua Bengio, published a joint open letter on Tuesday, calling for AI laboratories to default to "physically isolating" models capable of cyberattacks, meaning cutting off their connection to external networks. The open letter stated that in the short term, strong sandbox mechanisms, including network disconnection, should be adopted to implement high levels of containment for AI models with cyberattack capabilities. As of Wednesday, the number of signatories had increased to 450, including Stuart Russell from the University of California, Berkeley, and Ron Rivest, co-inventor of the RSA encryption system from the Massachusetts Institute of Technology.Before the release of the open letter, AI models experienced a summer of frequent "jailbreaking." In July, OpenAI admitted that GPT-5.6 Sol and an unpublished stronger model had breached closed testing environments, infiltrating the platform Hugging Face, which hosts over a million public AI models; these models were then being tested on the ExploitGym benchmark, which provides 898 real software vulnerabilities, requiring AI agents to convert each vulnerability into a usable attack. Subsequently, reports emerged that OpenAI models had unauthorized access to Australian government systems and obtained information from U.S. federal agencies.

According to Decrypt, the controversial promoter Sahil Arora, who launched a series of celebrity meme coins in 2024, is back at it again. His latest issuance, the Floyd Money Mayweather token, has been launched on the Solana trading platform Fomo, and Mayweather's social media account once promoted it. The token's market cap peaked at $1 million, but then plummeted 96% within three hours on Wednesday. According to information from various social media accounts, before Mayweather's account deleted the promotional post, the deployer's wallet had only cashed out about $5,000 in profit, and the token subsequently surged to a market cap peak of $1 million.Arora told Decrypt that the entire scheme was carefully planned, and the narrative of Mayweather missing out on profits actually drove up the token price. Fomo integrates social media elements, allowing users to see who is buying and selling, making it easy to create viral spread. Arora claims to have made seven figures in profit on the Mayweather token through other wallets, but on-chain analysis firm Bubblemaps believes this is unrealistic: the largest associated wallet cluster made about $105,000 in profit, and the top 100 traders collectively made only $306,000 in profit.Arora stated that he will replicate all his operations from 2024 on Pump.fun on Fomo, but with better strategies to allow participants to profit as well. He claims to have close relationships with all collaborating celebrities, making it very easy to persuade them to participate.

The Trump administration plans to transfer three senior officials from the U.S. artificial intelligence model testing agency CAISSI to the White House, reporting directly to White House technology policy director Michael Kratsios and national cyber director Sean Cairncross. The government had previously considered moving the entire agency to the White House but ultimately only adjusted part of the senior leadership; the related transfers have not yet been completed, and the identities of the three officials have not been disclosed. Since its establishment by the Biden administration, the agency has used three different names in less than two years, and Chris Fall, the head, resigned just three months after taking office in July of this year, with a successor yet to be determined.CAISSI is affiliated with the National Institute of Standards and Technology (NIST) under the U.S. Department of Commerce, has about 40 employees, and an annual budget of approximately $15 million, responsible for model safety assessments. According to informed sources, the agency is collaborating with the National Security Agency to test OpenAI's Astra and Anthropic's Mythos before public release; it has previously assessed DeepSeek V4 Pro and Zhiyuan GLM-5.3. Google confirmed that Gemini 4 Argon has participated in the voluntary testing process before the U.S. government release. The White House hopes that this adjustment will allow model testing experts to be more directly involved in policy decision-making.

ByteDance Seed recently sent an internal email announcing the establishment of the Seed Model-Alignment & Preparedness department, which will be led by Yan Lin and report to Seed head Wu Yonghui. Members of the original Responsible AI team and some members of the Chat team will transition to the new department.The department aims to strengthen investments in safety and alignment. The core of alignment is to ensure that model behavior is consistent with human intentions and safety constraints, in order to enhance the safety capabilities of Agent Models and prevent risks from cutting-edge capabilities such as chemical, biological, radiological, and nuclear threats.

According to The Information, a startup founded by Jeffrey Katzenberg and others is in talks with OpenAI to acquire the rights to use Sora, in order to adjust its features and sell related services.

On the eve of Double 11, the domestic large model manufacturer MiniMax officially settled in Tmall, opening its first brand official flagship store. The store has launched the M Plan subscription package, divided into three tiers: Go, Explore, and Build, supporting monthly and annual purchases.From now until October 14, consumers who purchase the above packages can enjoy a 50% discount on their first order for a limited time. Before this year's Double 11, Zhipu, Alibaba Cloud, and Qianwen Office have also opened stores on Tmall. Tmall has also launched the Token Recharge Center AI Space Station, consolidating multiple large model subscription services into a single entry point.

The artificial intelligence company Zhipu released a legal-specific model GLM-Legal, which is further trained based on GLM-5.3-Flash. In the Harvey LAB evaluation covering 24 categories of legal services, the task pass rate increased from 6.7% to 19.2%, nearly tripling the original version, surpassing models like Kimi K3 and Claude Fable 5, and ranking seventh on the Vals AI leaderboard.When reviewing software service contracts, GLM-Legal suggested shortening the termination notice period from 60 days to 30 days and required service providers to offer up to 90 days of handover support. In another software project dispute involving 40 documents, it found that the company reported to the client that the project was 60% complete, while internal records indicated only about 41%. It also identified emails from the CEO requesting to conceal risks and continue charging, and analyzed potential legal liabilities in conjunction with contract terms.According to independent evaluations by Vals AI, the cost for GLM-Legal to complete a test is $0.15, while the closely performing Gemini 4 Argon costs $10.55, a cost difference of about 70 times. GLM-Legal took an average of 72 minutes, while the latter took 44 minutes. Zhipu stated that the model was trained specifically on legal data and conducted reinforcement learning in tens of thousands of simulated legal environments. Currently, GLM-Legal requires an API trial application.

On October 12, 2026, William Zhou and Daniel Freeman, frontier red team researchers at the artificial intelligence company Anthropic, published a study. The article states that small models may have the ability to self-replicate on commercial IoT devices such as smart ovens and generic Android TV boxes within months. These replicas do not have a single shutdown point, and the doubling time may be measured in days, potentially adapting to unfamiliar devices on their own.Evaluation shows that a 9B model completed inference on 8 out of 9 hardware types. Models of 30B and below have been able to complete many single steps of replication, but errors occur when combined end-to-end; in simplified evaluations, GLM-5.2 and stronger models can achieve end-to-end replication. The team also distributed Gemma4-26B-A4B across 15 generic Android TV boxes, each costing about $55, operating at a speed of approximately 0.39 tokens/second, with each box holding two layers of the model.End-to-end testing was conducted on 8 GCP nodes behind a firewall, propagating through an unauthorized ADB port and setting up Llama 3.1-8B inference. Among the tested open-source models, only GLM-5.2 completed replication fully, transmitting to all 8 nodes in 4 out of 5 trials. The article divides autonomous replication and adaptation into four steps: discovering the target, obtaining execution, establishing inference, and continuing propagation from the new host, stating that there are currently no complete uncontrolled replicas.

Intel CEO Chen Liwu stated on October 12 in Singapore that the token costs of cutting-edge artificial intelligence models in China are lower, and the adoption speed is "very fast." At a seminar organized by DBS Bank, he mentioned that Chinese models like DeepSeek, Kimi, and GLM perform "well enough."Chen Liwu said, "I firmly believe in open source; it is the best way to drive efficiency." He added that the market will determine which models are best suited for different use cases. He also mentioned that quantum computing could become mainstream in the next 3 to 5 years, and software will play a role in driving the efficient use of some necessary infrastructure.