A chat response can be useful and still create more work: the user must pull material out of a scrolling exchange, move it into a document or design tool, and then explain the revision in a new prompt. IT Home and Leiphone reported that Doubao Work added an infinite canvas to task mode.
Enterprise AgentOS Separates Operations From the Model openJiuwen has put an enterprise-focused AgentOS into public code, giving teams that want to move beyond an agent demonstration a concrete stack to install and examine.
The meaningful part of XPENG’s latest robotaxi move is the invitation code. By opening online registration for XPENG YOYO, the automaker has begun taking a bounded ride-service test beyond its own employees.
Xiaomi has disclosed the working scale behind MiMo-V2.6’s reinforcement-learning training, giving builders a clearer sense of what sits behind long-horizon agent claims: not simply a model run, but a system that must generate, train, and grade vast amounts of work. The immediate consequence is practical.
A TianxiCode CLI v0.1.486 plus Deepseek-v4.1-Flash configuration, submitted by Tianxi AI Technology Platform, Lenovo, has taken the current top spot on the SWE-bench-Live Lite leaderboard, resolving 213 of 300 Python tasks.
A prompt that appears to be headed for Claude may pass through another service first, and that route can determine whether the person who sent it knows who can read or retain the exchange. In its threat-intelligence report, Anthropic alleges that some paths into its model used fake accounts, stolen credentials, and proxy services.
RobotEra has put VPP2 first on RoboDojo-Sim’s non-agent table, a result that gives robot builders a concrete new policy to inspect but not a reason to assume reliable physical task completion. VPP2’s 39.26 average Score and 32.26% success rate lead that table.
ByteDance Seed researchers report that retrieval outcomes varied depending on where information fell relative to a compression boundary in controlled synthetic tests. For anyone using long-context retrieval, that means a reassuring average score can still conceal a placement-dependent miss.
DiffuSpace has reportedly raised enough new capital to turn an unusual language-model design into a practical test for builders: can a system that revises many tokens in parallel make agent work more useful than the familiar, one-token-at-a-time approach? The Shenzhen company is the actor at the center of that question.
China has given local officials and companies more official routes to test AI in industry, but it has also made indiscriminate project-building more politically risky. The central guidance couples an AI push with a warning that chasing novelty or scale without discipline can bring accountability after major losses.
