AITOP WeeklyIssue 52351 stories

Meituan's LongCat-2.5, DeepSeek Ascend Kernels, and the Agent Sandbox Breach

Singapore·28 Sept – 4 Oct·Ten minutes

Chinese engineering took centre stage this week. Meituan shipped LongCat-2.5-Preview, a 1.6T-parameter mixture-of-experts model with roughly 48B active parameters and native 1M-token multimodal context. DeepSeek open-sourced six Huawei Ascend projects, with TileLang matrix kernels reaching 99.8% of chip peak speed, while Huawei's Grid-Interactive AIDC 1.0 delivered 18MW of power at SenseTime's site in just 45 days. Meanwhile Anthropic reported that Zhipu's GLM-5.3 approaches Claude Mythos in writing cyberattacks, and Nvidia responded with its Open Agent Safety Platform. For readers in Singapore, the message is that AI infrastructure choices are diversifying — and that sandboxing and audit trails must precede any agent deployment.

Five things this week

Enough to retell

Meituan Ships LongCat-2.5-Preview: 1.6T MoE With 1M Context

A food-delivery giant shipping a frontier-scale agent model shows how far Chinese consumer tech firms now push into foundation models, though the lack of published benchmarks makes real capability hard to judge. Developers building long-context or agentic workflows should track when independent evals and API access appear, since LongCat competes with open-weight MoE releases from Alibaba's Qwen and DeepSeek. Its 1M-token window targets codebase-scale agent tasks directly.

pandaily

DeepSeek open-sources six Huawei projects

This move significantly reduces dependency on Nvidia chips for AI training, potentially reshaping hardware competition in China. Developers can now optimize models for Ascend chips using DeepSeek's tools, though the licensing status of some projects limits immediate adoption. The timing coincides with Huawei's expected delivery of training chips this quarter.

rohanpaul_ai

Huawei Grid-Interactive AIDC 1.0 System

This system represents a significant advancement in energy-efficient AI infrastructure by directly linking power management with liquid cooling technology. For data center operators in Asia, particularly those using Ascend servers, this solution offers a practical approach to reducing energy consumption while maintaining high performance, potentially lowering operational costs by up to 30% compared to traditional cooling methods.

pandaily

Claude Sonnet 5.5 ranks third in Agent Arena

The performance-cost tradeoff reveals Anthropic's strategic positioning between premium and efficient models. Developers evaluating cost-performance ratios can now compare Sonnet 5.5's $2.74 per task against Opus 5.5's $1.58, with both leading different categories.

lmarena.ai

Multi-harness RL improves agent performance

This research reveals that agent performance varies dramatically across different testing environments, making multi-harness training crucial for robust AI agents. The open-source approach democratizes agent development, allowing anyone to train models that perform consistently across multiple harnesses rather than excelling in only one.

Hugging Face

US vs China

The same week, both sides

Dimension	United States	China	What it means for ASEAN
New models	Google released Gemini 4 Argon for government users, with 1M-token output, restricted access.	Meituan's LongCat-2.5-Preview: 1.6T MoE, ~48B active, native 1M multimodal context.	Long context and agentic use are bets on both sides.
Prices	Claude Sonnet 5.5 costs $2.74 median per Agent Arena task, +12.5% net improvement.	Alibaba Qwen3.8-27B: $0.45 per ARC-AGI v2 task, just $0.22 on v1.	Chinese models remain far cheaper per task.
Benchmarks	NVIDIA research: accuracy drops 62.8% across seven open models as context grows 4K to 128K.	Qwen3.8-27B scores 87.5% on ARC-AGI v1 but 42.4% on v2.	Long-context agent reliability is a shared weakness.
Policy	Nvidia launched Open Agent Safety Platform, sandboxing agents and restricting sensitive data by default.	Anthropic says Zhipu GLM-5.3 nears Claude Mythos in cyberattack capability; unlocked builds circulating.	ASEAN firms should deploy sandboxing and audits before agents.