agent

BOSS Zhipin

2026-10-26 03:57:32

BOSS Zhipin’s Nanbeige Lab releases open-source Nanbeige4.3-33B for coding and office agents

Nanbeige Lab, a research unit under recruitment platform BOSS Zhipin, has released Nanbeige4.3-33B as an open-source model aimed at coding and office agents. The model supports a 512K context window and is distributed under the Apache 2.0 license. According to the official comparison, it was benchmarked against Qwen3.8-27B, Meta Muse Glimmer 30B, and Gemma 4 31B, and ranked first in 17 of 20 disclosed tests. In Terminal-Bench 2.1, Nanbeige4.3-33B scored 80.9 versus Qwen’s 73.0. On DeepSWE, it posted 65.5 compared with 42.5, while its score on the office automation benchmark Automation Bench reached 54.2, above Qwen’s 38.3. Qwen remained ahead on SWE-Bench Pro and GDPval. The lab said it had previously focused on 3B-scale small models. Its earlier Nanbeige4.2-3B improved effective compute depth by reusing Transformer layers without increasing parameter count. This time, the team trained a 33B model first and plans to use it as a teacher model, distilling coding and tool-use capabilities into the upcoming Nanbeige4.3-3B.

120

Mojo

2026-10-22 02:42:39

Mojo lands on the App Store with iPhone support for creating trading agents

AI trading platform Mojo said its app is now available on Apple’s App Store, opening iPhone access for users who want to turn trading ideas into agents inside the mobile app. According to Mojo, users can create an agent on iPhone, review the proposed plan, approve it, and then monitor the agent’s activity from within the app. The company also issued migration instructions for existing users coming from the TestFlight version. Those users need to remove the beta app first, install the new public release version 1.0.3 or later, sign in with the same login method they used before, and check their account details after logging in. Mojo said new iOS subscriptions are now handled through Apple’s in-app purchase system. Existing members were told to confirm their current access status before subscribing again, so they do not end up paying twice for the same service.

410

Shengshu Tech

2026-10-21 13:01:42

Shengshu Technology starts public livestream of Motubrain 2.0 training process

Shengshu Technology’s world model team has started a public livestream of the full Motubrain 2.0 training process on Oct. 21, Beijing time, with the broadcast scheduled to run for half a month. According to Techub News, the stream is designed to show the public how an embodied model is trained and how the research progresses over time. The livestream is set to cover four parts of the workflow: pre-training, post-training, Agent task execution, and exploration of recursive self-improvement, or RSI. The team said the goal is to present the full path from training and testing to capability iteration, rather than only showing final results. Motubrain is Shengshu Technology’s embodied model series, built to provide robots with a general-purpose “brain.” The team said the open broadcast is meant to open up the “black box” of embodied model training, giving researchers and developers a closer look at concrete issues that arise during training and how those issues are addressed. It also aims to share experience with technical peers and discuss topics including large-scale training and cross-task generalization.

410

Today AI

2026-10-21 03:52:45

Today AI says it has raised more than $200 million and started training its own model

Today AI, founded by Teambition founder and former Doubao PC lead Qi Junyuan, said it has raised more than $200 million in total. Backers named by the company include IDG Capital, Tencent, 5Y Capital, Ant Group, Sequoia China, Transsion, Tongge Venture Capital and Shanghai Jiao Tong University Hanyuan Asset. The company had originally planned to build software first, then hardware, and only later train its own model. It has now changed that sequence and is pushing software and model development at the same time, with an early experimental model already in place. According to the company, current large models are getting better at coding, but their ability to communicate naturally with users and understand personal preferences has not improved at the same pace. Today AI said it wants to train a model tailored for personal assistant use cases, while hardware will be developed with partners. The team also acknowledged that the product still falls short of its vision for a personal AI assistant, citing limits in proactive understanding, adaptation and task judgment.

520

OpenAI

2026-10-19 01:33:19

How Codex went from a failed cloud debut to the shell beneath ChatGPT desktop

OpenAI’s Codex has emerged as one of the company’s most consequential product stories, and a long-form report by Every CEO Dan Shipper lays out how that happened in detail. Based on interviews with OpenAI co-founder Greg Brockman and six core Codex team members, the report traces Codex from an early cloud-first launch that struggled to gain traction to a local-agent pivot, a rebuilt core architecture, and eventually a desktop app that became the foundation for ChatGPT’s desktop experience. The account points to several turning points. The first Codex Cloud, released in May 2025 after roughly seven weeks of work, ran into two problems at once: developers faced difficult real-world environment setup, and the model itself was not yet strong enough to handle complex tasks independently. OpenAI then revived a local path it had previously sidelined, rebuilt the underlying agent in Rust, and paired it with stronger post-trained GPT-5 coding models. By October 2025, Codex usage had grown more than 10x, with IDE extensions overtaking the terminal as the main entry point. The next shift was even bolder. Rather than doubling down on the terminal and IDE alone, the team built a native desktop app designed to manage multiple agents at once, and deliberately left out a code editor. That bet only worked after newer Codex models improved. The desktop app launched on Feb. 2, 2026, reached more than 2 million weekly active users in March and more than 5 million by June, and later merged with the ChatGPT desktop experience in July.

580

ChainCatcher

2026-10-19 00:54:49

CITIC Securities says Harness could be a key link in DeepSeek’s platform push

According to a report cited by Jin10 and carried by ChainCatcher, CITIC Securities said Harness could become a key part of DeepSeek’s shift from supplying models to building a broader platform ecosystem. The brokerage said DeepSeek may turn its open-source and engineering strengths into user growth, ecosystem influence, and commercial capability. As competition in the Agent segment moves beyond model performance toward stable and easy task delivery, the report said DeepSeek is expanding user and task coverage through plug-in architecture, preset scenarios, and cross-Agent collaboration. CITIC Securities also said the company’s longer-term value lies in ecosystem expansion, accumulated demand, stronger product stickiness from configuration and workflow buildup, and traceability features that support security governance. On top of official model access, the report said the company could also explore platform services and ecosystem transactions to diversify revenue sources and support reinvestment into products and the ecosystem. CITIC Securities said it remains optimistic about the continued rise of China’s domestic large language models.

530

Liquid AI

2026-10-14 08:00:18

Liquid AI updates 24B local model, posts higher Agent scores than larger Qwen model

Liquid AI has introduced LFM2.5-24B-A2B, a new local model aimed at improving on-device Agent reasoning and tool use. The model keeps the same 24 billion total parameters as its predecessor while activating only about 2 billion parameters per run. Compared with LFM2-24B-A2B, the new release expands pretraining data from 17 trillion to 33 trillion tokens and increases context length from 32K to 131K. Liquid AI said the model is trained to reason before answering and was reinforced directly inside Agent frameworks including OpenClaw, Hermes Agent, Pi, and OpenCode to practice multi-step tool calling and result handling. In the company’s demo, Hermes Agent ran on a MacBook without using a cloud model API, searched trending Hugging Face projects, compiled a briefing, and continued after follow-up prompts. Liquid AI’s tests said the model generated 184 tokens per second on Apple’s M5 Max, with peak speed reaching 1.7x that of comparable MoE models. On Agent benchmarks, it scored 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, and 27.69 on AutomationBench, more than three times Qwen’s result. On SWE-bench Verified, however, it scored 51%, below Qwen’s 66.2%.

1000

AI Deng

2026-10-13 02:21:01

‘AI Deng’ Goes Viral in China as AI Talk Turns Into Office Theater and Social Posturing

A commentary article published by WeChat account New Weekly describes the rise of “AI Deng,” a mocking label for people who constantly talk about prompts, agents and multimodal models while producing little practical output. The piece argues that the phenomenon has spread across offices, social media feeds, conferences and executive culture, where AI is presented less as a tool and more as a badge of status. Employees are pushed into AI training, AI-generated reports and endless revisions, while many of the promised gains remain vague or unproven. The article also points to a widely discussed data post from an overseas solo AI entrepreneur earlier this year. While the numbers and methodology were disputed, the post tapped into a broader anxiety: the world may look surrounded by AI largely because a small minority speaks the loudest. In the version cited by the article, about 6.8 billion people had never used AI, 1.3 billion had used free conversational AI, and only 0.04% had used coding or agent tools. By linking the term to the older Chinese internet concept of “laodeng,” the article says the issue is not age or the technology itself, but a style of behavior rooted in hierarchy, performance and the need to appear ahead of the curve. It frames “AI Deng” as a human problem shaped by status, anxiety and power rather than a problem caused by AI alone.

760