agent

Tencent

2026-09-02 07:00:53

Tencent’s WorkBuddy launches open platform with more than 100 ecosystem partners

Tencent’s WorkBuddy has launched its open platform and brought in more than 100 ecosystem partners at the first stage. The company said the platform will open up its underlying Agent capabilities to smart hardware makers, industry application providers, and developers. At the launch event, nine co-branded smart devices were unveiled, with partners including Plaud, Rokid, Insta360, iFlytek, Anker, Monmo, and JD Smart. More than 30 industry applications were also connected, spanning finance, legal services, healthcare, education, and public welfare across more than 20 sectors. WorkBuddy had already worked with Liweike Technology on the X-AI memory glasses and reached an AI glasses ecosystem partnership with Rokid.

1420

Alibaba

2026-09-02 02:56:43

Alibaba's Qwen3.8-Max Gets 0902 Upgrade: Coding Scores Surge, Beats Claude Opus 5 in Three Tests

Alibaba's Qwen team has released Qwen3.8-Max-0902, an upgraded version of its flagship model just one month after the original launch. Parameter count stays at 2.4T with 1M token context. The update focuses on coding and agent capabilities, with all 8 coding benchmarks showing improvement. The new model surpasses Claude Opus 5 in MLS-Bench-Lite, SWE-Atlas QnA, and QwenSWEbench V2. Multimodal scores increased modestly by 0.4-3 points. The API is now available at unchanged pricing: $2 per million input tokens and $6 per million output tokens.

1200

PONS

2026-09-01 12:28:29

PONS Market Cap Dips Below $400M, B.AI Token Throughput Exceeds 1.1 Trillion Daily

PONS market cap briefly fell below $400 million. Meanwhile, B.AI's daily token throughput exceeded 1.1 trillion, signaling accelerating scale effects in AI infrastructure. On the Agent side, B.AI continues advancing x402 payments, on-chain identity, Agent Wallet, and MCP/Skills capabilities, enabling agents to autonomously handle payments, settlements, and on-chain operations.

1480

B.AI

2026-09-01 12:21:41

B.AI says daily token throughput topped 1.1 trillion, reaching trillion-scale usage in 145 days

B.AI, an AI infrastructure platform, said its daily token throughput has exceeded 1.1 trillion for the first time, a milestone reached roughly five months after launch. The company said its user base has passed 2.28 million and that daily token processing rose from 21.1 million on April 9 to more than 1.1 trillion in 145 days. In its comparison, B.AI said OpenRouter took about 2 years and 9 months to reach a similar daily throughput level, putting B.AI’s pace at roughly seven times faster based on that public timeline. The article also said that from Aug. 17 to Aug. 31 alone, B.AI’s daily throughput grew about 242-fold, while cumulative processed volume hit 6.86 trillion tokens and the platform added more than 196,000 API users. B.AI framed the growth as evidence that it is moving beyond a unified model access point toward routing, metering, payments, settlement and Agent infrastructure, a direction it describes as building an “intelligent global settlement layer.” The platform currently supports domestic model providers including DeepSeek, GLM, Qwen, Kimi, MiniMax, Tencent Hunyuan and Xiaomi MiMo, while also offering traditional payment methods and crypto top-ups across TRON, BNB Chain and Ethereum.

1070

X-Agent

2026-09-01 04:09:23

X-Agent AI MCP Hackathon 2026 to open Sept. 2 with two tracks and prizes in USDT and X-Points

X-Agent said its X-Agent AI MCP Hackathon 2026 will officially begin on Sept. 2, inviting developers and teams worldwide to submit Agent and MCP applications with practical use cases and verifiable functionality. The online global event is supported by OlaXBT and includes two separate competition tracks. The open innovation track calls for practical API-driven Agent or MCP capabilities built around AI, crypto assets, data, automation, and Agent infrastructure. A second track, the OlaXBT × X-Agent Trading Challenge, gives participating teams access to OlaXBT Nexus MCP tools for strategy development, backtesting, performance analysis, and market data, allowing them to validate trading strategies and build related Agent or MCP applications. Prize pools include USDT and X-Points. The first-place team in each track will receive 500 USDT, while the top five teams in each track will share X-Points. According to the announcement, X-Points can be used to take part in the redemption for X-Agent airdrop token $XAGT. Registration and development run from Sept. 2 to Sept. 19, judging is scheduled for Sept. 20 through Oct. 1, and winners are expected to be announced between Oct. 2 and Oct. 4.

1740

DeepSeek

2026-08-31 11:03:53

DeepSeek open-sources first vision model in the V4 lineup

DeepSeek has open-sourced DeepSeek-V4-Flash-Vision-Exp, the first experimental multimodal model released under its DeepSeek-V4 series. The model is built on the V4-Flash architecture and adds a vision module, with image understanding capabilities obtained through continued training. According to the brief, it can interpret screenshots, read charts, and use tools to complete agent tasks. On pure text-based agent tasks, the model is said to maintain performance comparable to V4-Flash. The release comes with 305B parameters and is distributed under the MIT license. The announcement frames the model as an experimental step for the V4 family rather than a general summary of the broader lineup.

1150

CREAO

2026-08-31 08:33:15

CREAO raises fresh strategic funding to build self-improving Agent harness systems

Enterprise Agent platform CREAO has closed a new strategic funding round worth tens of millions of dollars, bringing its cumulative financing to more than $30 million. The company is now focused on self-iterating harness systems, a control layer outside the model that governs how an Agent plans tasks, calls tools, uses memory, selects models, and handles failures. CREAO said it has already put part of that automated correction loop into operation: after an Agent completes a task, the system checks the output, opens a ticket if a problem is found, and then uses AI to investigate the cause, write a fix, and verify the result. New versions are first tested on a small amount of real traffic, with automatic rollback if performance gets worse. The company had previously disclosed this system publicly. CREAO’s next step is to extend those improvements to more parts of the harness so repeated failures in the same tool or workflow can lead to broader execution changes that newly created Agents can reuse. It also plans to move some of that accumulated experience into model training through vertical post-training on open-source models, using lower-cost specialized models for common tasks and frontier models for more complex ones.

1360

BlockBeats

2026-08-30 07:16:34

Reflexion Co-Author Says Instinct’s Edge May Be Its Handling of Long-Term User State

Ashwin Gopinath, a former MIT professor and co-author of Reflexion, says the fast-rising personal agent Instinct may stand out not because it executes tasks better than rivals or stores ordinary long-term memory, but because it appears to maintain a user’s long-term state over time. He said the view comes from hands-on use, not internal access. In his assessment, Instinct does not look dramatically stronger than Hermes or OpenClaw on raw task execution. What feels different is that the agent seems to understand the user better the longer it is used, without becoming bogged down by an ever-growing history of past interactions. Gopinath thinks the key may be that Instinct does not simply archive every chat and action log and then retrieve them with vector search or a knowledge graph. Instead, it keeps an evolving picture of what still matters: which preferences remain valid, which facts have changed, which tasks are unfinished, and whether the next move should be to wait, follow up, or act. He describes the process as “memory compilation,” where useful signals are filtered out of messy life records and compressed into a continuously updated state. For long-running agents, he argues, the hard problem is no longer just remembering more, but knowing what should persist, what should expire, and what change should trigger action.

1610