agent

Alibaba

2026-09-22 08:38:57

Alibaba launches Qwen Intelligence as a system-level agent stack for phone makers

Alibaba has introduced Qwen Intelligence, a mobile-focused AI offering aimed at smartphone manufacturers rather than end users. The product is not a standalone app and does not involve Alibaba building its own handset. Instead, it packages the Qwen model family, an agent platform, and mobile scenario capabilities for OEM integration. The initial release includes three components: Mobile Planner Agent for task planning, Mobile-Use Agent for operating the phone, and Mobile Creative Agent for image generation. Alibaba said the planner can break down complex tasks, arrange execution order, and retain both working memory and long-term memory to keep track of context and user preferences. The operation agent prioritizes MCP, API, and DeepLink access when available, and falls back to visual recognition and GUI interaction when direct interfaces are unavailable. For high-risk actions such as payments or data deletion, final confirmation remains with the user. Alibaba also said the creative agent can turn a single prompt into an image by decomposing the workflow and calling different capabilities. The company said the process has been reduced from 100 steps to 8 steps, with the first image generated in about 3 seconds. The system has already been deployed with Honor MagicOS, and the two sides said complex long-horizon tasks can run for more than 100 consecutive steps, with an end-to-end service loop closure rate of 90%.

820

B.AI

2026-09-19 12:14:22

B.AI says registered users topped 2.7 million as it pushes an AGI economy network

B.AI said on Sept. 19 that its cumulative registered user base reached 2,700,048, rising from 2.6 million to above 2.7 million. The platform attributed the increase to continued optimization of model supply and pricing, along with multiple discounts and user benefits that it said have lowered the cost of accessing AI tools. B.AI also said it is expanding the functional scope of Agents, moving them beyond task execution into payments, collaboration, and value creation. According to the platform, each user milestone reflects more users and Agents joining its intelligent network, with value connections inside what it described as an AGI economy network becoming more active. The company also directed users to b.ai to explore additional frontier models and AI capabilities. The update was carried by Odaily.

840

DeepSeek

2026-09-15 06:57:45

DeepSeek engineer says AI may soon catch up in kernel development work

DeepSeek kernel engineer Liu Shengyu said AI is rapidly closing in on parts of his job, including low-level operator development and optimization. In a recent post, he said he wrote the main Attention kernel for DeepSeek V4.1, but at the current pace of progress, AI will likely match or exceed his kernel-writing ability within six months to a year. He said that a year ago, AI could mainly help with documentation lookup, code reading, and bug finding. Now it can read CUDA, PTX, and SASS, analyze instruction stalls, and optimize kernels on its own. Liu said the next step may be AI systems that design scheduling plans, evaluate performance, and complete implementation independently. He added that he is unlikely to lose his job outright, but may be forced to shift roles, moving from writing kernels by hand to operating AI agents. He also explained why he remains at DeepSeek, saying frontier AI should be offered openly and cheaply to everyone, and warning against a future in which the most powerful AI is controlled by a small number of companies.

910

Feishu

2026-09-15 06:08:15

Feishu 8.0 turns Doubao Work Partner into a team-wide Agent

Feishu has released version 8.0 and upgraded "Doubao Work Partner" into an Agent designed for entire teams rather than individual users. The Agent has its own identity, permissions, and memory, and it can join Feishu group chats directly. After a company completes a unified setup, employees can work with the same Agent instead of configuring one separately for each person. Within the scope of granted authorization, the Agent can read documents, spreadsheets, meetings, calendars, and enterprise business systems. It can also directly operate documents, multidimensional spreadsheets, schedules, meetings, cloud storage, and approval workflows. In addition to syncing information, reviewing documents, and organizing weekly reports, it can proactively carry out tasks based on work progress. Liang Rubo also addressed an earlier organizational adjustment involving Feishu, Doubao, and Volcano Engine.

1070

AI assistants

2026-09-15 04:14:41

Assistant Benchmark launches real-world leaderboard for personal AI agents, with Muse on top for now

Assistant Benchmark, an independent evaluation project, has started comparing personal AI assistants through real-world tasks rather than abstract benchmarks. The test setup asks agents to complete practical assignments such as booking hotels, choosing restaurants, shopping, replying to emails, and connecting third-party services. Scores are based on whether those tasks are actually completed, and each scoring category is tied to a public fixed task that must include a real execution record before a score is awarded. At the current stage, Muse ranks first with an average score of 9.1 across seven completed categories. Instinct has been tested across 11 categories and holds an average of 8.4, while Grok Bot has completed seven categories with an average of 7.3. Muse posted perfect 10-point results in shopping, email, and third-party app connection tasks. The benchmark should be read as a live, evolving experience board rather than a final ranking. Muse’s 9.1 only reflects the average from seven tested categories, not a full result, and each category is still being measured with only one fixed task for now. As more follow-up testing is added, both the scores and the ranking may change.

980

Tencent

2026-09-14 10:55:00

Tencent’s WeKnora adds long-term memory, tool workspace and code execution in v0.8.0

Tencent’s WeChat team has expanded its open-source knowledge framework WeKnora with a new v0.8.0 release that adds cross-session long-term memory and rounds out its Skills and code execution stack. WeKnora was originally built to turn enterprise documents into retrieval-augmented generation, or RAG, knowledge bases that support search and question answering. Over time, the framework added Agent, Model Context Protocol, or MCP, and Skills support. In the latest update, WeKnora can store user profile details, preferences, facts and tasks, then call that information back in later conversations. The system does not activate automatically extracted memories by default; users must confirm them first. Skills are now packaged as standalone workspace resources and can be installed from ClawHub, SkillHub, GitHub, GitLab or zip files. During task execution, the agent can also enter Docker, E2B or Cube sandboxes to run commands, read and write files, and send generated files back to the chat window. Admins can restrict which networks those sandboxes may access.

1440

Tencent

2026-09-11 10:20:30

Tencent unveils T1 for long-horizon terminal Agent tasks, with 300-plus tool calls per run

Tencent has introduced T1, a terminal-focused Agent model trained from Qwen3.5-122B-A10B for complex Linux command-line tasks. The model can call tools for more than 300 rounds within a single task, targeting longer and more demanding terminal workflows. On Terminal-Bench 2.1, T1 scored 64.0%, up from the base model’s 43.8%, a gain of 20.2 points. Tencent said most of that improvement came from reinforcement learning. Supervised fine-tuning, or SFT, lifted the score to 49.4%, and reinforcement learning added another 14.6 points after that. The team also prepared about 15,000 terminal tasks, each paired with automated tests. During training, the Agent could still receive rewards for completing only part of a task’s requirements rather than the full assignment. Tencent also described a training-inference mismatch in long tasks: the same content may be tokenized differently between execution and later training, while a mixture-of-experts model may route the sequence to a different set of experts. To address that, T1 records the generated tokens and the experts selected at the time of execution, then replays them during training. Tencent said this reduced the mismatch by about one-third.

1140

Shi Kaiwen

2026-09-10 09:46:44

Former Feishu VP Shi Kaiwen drops startup plan and joins a large-model company

BlockBeats reported that Shi Kaiwen, former vice president of product at Feishu, has abandoned his device-side Agent startup and joined a large-model company as a product partner. The move came less than three months after he announced the venture in late June. Shi had originally planned to build a device-side Agent, with team members from ByteDance and Shanghai Jiao Tong University. A June report had said that Sequoia China and Jinqiu Fund took part in the project’s early financing, but the funds had not actually been wired. The startup also had not publicly disclosed a concrete product form. Shi joined ByteDance in 2019, worked on Feishu’s product direction for years, and led the Feishu Bitable product.

1010