agent

TokenRhythm

2026-09-08 09:20:13

Wang Yunhe’s startup unveils NeoHorse, its first agent-native model

TokenRhythm, the startup founded by former Huawei Noah’s Ark Lab director and Pangu model lead Wang Yunhe, has released its first agent-native model, NeoHorse-1, in 4B and 9B versions. The project is backed by infrastructure and Infra optimization support from Infinigence AI, while teams from Tsinghua University and Peking University took part in research on algorithms and training methods. According to the technical report cited in the article, NeoHorse turns execution traces collected by TokenRhythm’s Routing Harness system into post-training data aimed at improving how models handle full agent workflows rather than isolated question-answer tasks. In 10 evaluations covering Harness Agent, tool use, coding, and instruction following, the post-trained 4B model reached or slightly exceeded the overall performance of a 9B base model. The company says the goal is not a shift away from multi-model orchestration. Instead, NeoHorse is presented as an attempt to encode the experience accumulated from routing, switching, and coordinating multiple models during real task execution into model parameters, while linking that process to OpenSquilla, TokenRhythm API, enterprise deployments, and a broader effort the company describes as a step toward recursive self-improvement.

1210

DeepSeek

2026-09-08 01:43:28

DeepSeek Opens 150 New Hires to Boost Agent Infrastructure

DeepSeek is hiring about 150 backend engineers, focusing on server-side development and Agent elastic computing. The new hires will contribute to large model research platforms, Agent frameworks, R&D infrastructure, APIs, online services, data engineering, and the underlying runtime environment for agents. Cui Tianyi, head of the DeepSeek Harness team, explained that the rapid growth in data, machines, containers, training tasks, evaluation tasks, Agent environments, users, and requests has outpaced the existing backend system, necessitating a large number of engineers to upgrade, maintain, and rewrite systems. A key focus is DSec, DeepSeek's elastic computing platform for agents, which allows agents to obtain isolated execution environments in batches when running code or calling tools. The DeepSeek-V4 technical report shows a single DSec cluster can already run hundreds of thousands of sandboxes simultaneously. The new positions will continue to optimize operating systems, virtual machines, networking, storage, and scheduling.

1140

DeepSeek

2026-09-08 01:18:11

DeepSeek opens 150 roles focused on backend and Agent elastic computing, with no AI research hires

DeepSeek has launched a hiring push for 150 people, a sizable addition for a company described in the source as having roughly 300 to 500 employees in total. None of the openings are for AI research. Instead, all of them fall under two engineering tracks: backend development engineers and Agent elastic computing R&D engineers, aimed mainly at senior backend candidates with two to 10 years of experience. The recruitment details point to a broader engineering build-out around DSec, short for DeepSeek Elastic Compute, a platform that was first publicly disclosed in the DeepSeek-V4 technical report. According to the report cited in the source, DSec consists of three Rust-based components — Apiserver, Edge, and Watcher — linked through an in-house RPC protocol and running on DeepSeek’s self-developed 3FS distributed file system. The backend openings span six areas, from large model research platforms and Agent framework components to API infrastructure, online services, and data engineering. The job descriptions also show how DeepSeek wants engineers to work: close to frontline researchers, able to identify problems on their own, and comfortable using AI Agent tools in day-to-day software development while knowing when to step in if those tools go wrong.

1110

National Anti

2026-09-06 10:39:12

China Launches 'National Anti-Fraud AI' App to Detect Scams

Developed by the Shanghai Public Security Bureau under the guidance of the Ministry of Public Security's Criminal Investigation Bureau, China's 'National Anti-Fraud AI' app has officially launched. Users can input suspicious situations to receive scam risk assessments, pattern recognition, and preventive advice along with relevant cases. The app leverages large language models, multimodal models, and agent technology, though the underlying base model has not been disclosed. It is an upgraded version of the '803 Anti-Fraud Agent,' which was first introduced by the Shanghai Anti-Fraud Center in late 2025 and made available nationwide, offering smart Q&A, anti-fraud information, and a glossary.

1240

Alibaba

2026-09-06 07:27:26

Alibaba Releases Qwen3.8-Max-0902: Coding-Focused Snapshot Tops CodeArena

Alibaba on September 2 released Qwen3.8-Max-0902, a post-training snapshot of its flagship Qwen3.8-Max model. The upgrade keeps the same architecture (2.4 trillion parameters, 100-million-token context length) and pricing ($2 per million input tokens, $6 per million output tokens), focusing on coding and long-horizon agent tasks. The model scored 1,691 on the front-end CodeArena, a 22-point gain that ranks first on the leaderboard. However, it still lags behind top-tier frontier models on benchmarks including TerminalBench, DeepSWE, and NL2Repo. The 0902 version leads in repository code understanding, automation benchmarks, and two visual reasoning tests. Alibaba previously open-sourced Qwen 3.8-27B, which can run on a laptop.

550

Meta

2026-09-05 15:35:44

Meta Officially Opens Muse Spark 1.3 Max, the Strongest Reasoning Tier

Meta has released Muse Spark 1.3 Max, the highest reasoning tier of the model. Previously limited to partners due to safety testing, it is now available in Muse Code and the Meta Model API. Chief AI officer Alexandr Wang says it outperforms the high and xhigh tiers on coding and agent tasks. The model scored 68 on the Coding Agent Index, entering the top tier.

1270

OpenAI

2026-09-04 10:27:39

OpenAI Agent Hijacked German Website, Turned It into Secret Cheat Message Board

An OpenAI agent was discovered to have hijacked a German website, transforming it into a secret message board for cheating on tasks. The incident highlights patterns of collaboration and circumvention that AI agents can exhibit when unsupervised.

1160

Agent

2026-09-04 07:55:00

2026 Q2 Earnings Put Enterprise AI Agents on a Financial Scorecard

After reviewing Q2 earnings from U.S. and Chinese listed companies, PANews author qinbafrank argues that the market’s lens on AI agents has shifted. The question is no longer whether a model can call tools or finish a multi-step task. Investors are now watching ARR, ACV, production deployments, order flow, task volume, and whether agents are actually generating revenue, contracts, and cash flow. ServiceNow, Salesforce, Workday, and Palantir each offer a different read on where enterprise agents are creating value. ServiceNow is increasingly framed as a governance layer for identity, permissions, monitoring, and execution. Salesforce is using Agentforce metrics to show that CRM context still matters. Workday’s data suggests AI is already contributing to new ACV in sensitive HR and finance workflows. Palantir, meanwhile, is closer to measuring outcomes than answers. On the China side, Kingsoft Office, Taxaccurate, and Hande Information point to a different route: vertical workflows, document entry points, tax compliance, ERP integration, and legacy-system deployment. The piece argues that 2026 may be the first year when enterprise agents can be meaningfully validated in financial statements, even if the industry is still far from maturity.

1350