 DeepSeek offers developers an open-source agent harness to compete with Anthropic's Claude CoworkDeepSeek provides developers with an open-source agent harness to compete against Anthropic's Claude Cowork, enabling the operation of AI agents through a web-based interface and plugin architecture; this is offered under an MIT license and, with a rapidly evolving preview version, supports developers in building flexible tools and innovating within a competitive ecosystem. |
AiPulse — Issue No. 28
COVER
Pulse
Apple, with Alibaba's support, is developing a custom large language model for the Chinese market and aims to provide on-device generative AI services, seeking a competitive edge and overcoming regulatory hurdles. |
GLM-5.3 has arrived with advanced cyber capabilities and may have discovered a significant security vulnerability in Cursor. |
Global Lens
Alibaba Cloud releases Qwen3.8-27B model weights under Apache 2.0 licenseThis 270-billion-parameter model can run on local GPUs and, in some benchmarks, surpasses Anthropic's Claude Opus 4.6 Max performance, offering commercial usage under an open license. |
Google Gemini 3.7 Flash accelerates enterprise automation by enhancing coding and multi-step workflow capabilities, while halving API costs until the end of 2026 to make AI integration more economical for companies. |
This mode aims to increase the AI model's response speed by 14 times, boosting efficiency in urgent business processes and high-volume data analysis. |
This implementation allows developers to optimize API usage costs based on peak hours, enhancing the economic efficiency of the service. |
Frontier
Dual-Flow Transformer activates additional computation only from the end of the text during response generationThis approach decouples the extra computation used in large language model response generation from the prefill stage, reducing total inference cost on hardware and providing finer control over prefilling and continuation costs, especially in MoE models. |
MARCH scales recurrent memory using content-routed state anchorsMARCH stores all past states in a memory bank and queries these anchors at each token with content-conditioned keys, increasing the memory capacity of fixed-size recurrent models in long contexts while keeping computational cost low; after pretraining, it consistently outperforms linear attention variants in joint reasoning, LongBench, and retrieval-in-context tasks. |
By summer 2026, Chinese labs have released larger open models than those from the US, yet download statistics indicate that value in the AI ecosystem still accumulates in low-parameter models; this situation reshapes the direction of the open model ecosystem and the strategies of hardware companies. |
Watch & Listen
 Türkçe altyazı7 AI Terms You Need to Know: Agents, RAG, ASI and MoreThe video explains the seven essential terms you need to know in the rapidly evolving world of artificial intelligence, particularly focusing on AI agents capable of making independent decisions. |
Fun & Curious
This achievement demonstrates that hydrogen can be used as a reliable fuel throughout all flight phases—from takeoff to cruise and landing—in a modern aero gas turbine, marking a critical step for low-carbon aviation. |
This design is powered by a chemical process that converts transuranic actinides extracted from waste fuel into chloride salts, enabling a fast neutron spectrum reactor that delivers 300-500 MW output and dissipates heat through natural convection without the need for emergency power. |
FOLLOW
Get new issues in your feed reader — add this to an app such as Feedly, Inoreader or NetNewsWire. Follow by RSS |