open-source

DeepSeek’s experimental V3.2-Exp model introduces “DeepSeek Sparse Attention” (DSA) to enhance efficiency, reduce costs, and handle longer documents. DSA filters less relevant data, potentially halving operational costs. While promising faster, cheaper AI deployment, concerns exist about potential loss of critical data nuances and impact on model reliability. Designed for Chinese AI chips, DeepSeek’s open-source approach encourages collaboration but faces potential patent challenges. The focus on efficiency positions DeepSeek competitively in the evolving AI landscape.

Huawei introduced SuperPoD at HUAWEI CONNECT 2025, a new AI infrastructure architecture that aggregates thousands of AI chips into a unified resource using UnifiedBus (UB). This creates a “supercomputer” from distributed servers, designed to address the limitations of traditional architectures. The Atlas 950 SuperPoD utilizes up to 8,192 Ascend 950DT chips, with future plans for the larger Atlas 960. Beyond AI, TaiShan 950 SuperPoD targets general-purpose computing. Huawei’s open-source approach with UnifiedBus 2.0 aims to accelerate innovation and foster broad industry participation in AI infrastructure development.

The UAE’s Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) introduced K2 Think, a compact reasoning model challenging OpenAI and DeepSeek. Developed with G42, K2 Think utilizes 32 billion parameters and Alibaba’s Qwen 2.5, achieving competitive benchmark results in math and science. MBZUAI emphasizes its focus on scientific advancements, aiming to accelerate research with limited resources. This development underscores the UAE’s ambition to become a key player in the global AI landscape, joining the US and China in the race for AI innovation.

The 2025 Open Atom Open Source Ecology Conference in Beijing spotlighted open source’s role in industry. Topway Info & Communication and Kaihong Zhigu showcased their contributions, receiving the “Open Source Pioneer Enterprise” award. They launched the OpenHarmony Co-construction Map 4.0 and participated in talent development initiatives. Kaihong Zhigu also joined the OpenHarmony AI Agent TSG, aiming to integrate AI with the operating system. They presented on “AI + Harmony” applications across education and transportation.

At the 2025 OpenAtom conference, Huawei’s Zhang Xiaofeng highlighted OpenHarmony’s progress in community growth (over 9,000 contributors), technological advancements, and industry adoption with over 1,200 compatible products. OpenHarmony supports diverse sectors, from finance to transportation. Customized OS distributions like “YiHong” are driving digital transformation. Integration of HarmonyOS in government and enterprise office applications is accelerating, promoting efficient office experiences. Zhang Xiaofeng called for increased collaboration to build a thriving OpenHarmony ecosystem.

Kimi’s K2 model launched with open-source architecture and disruptive pricing, claiming superior code generation to Claude Sonnet4. Dangbei AI, a Chinese AI platform, is among the first to integrate K2 into its “super AI toolbox,” complementing DeepSeek R1-0528 and Doubao Large Model 1.6. K2’s strengths include its accessible architecture and low cost (4 yuan per million tokens). Dangbei AI’s integration strategy focuses on comprehensive scenario coverage with seamless model switching and end-to-end solutions, alongside features like no-login access and incognito mode.

Baidu has open-sourced its ERNIE 4.5 AI models, featuring multi-modal Mixture-of-Experts with enhanced orthogonality. Modong Space, China’s national computing platform, is the first to integrate these advanced models. ERNIE 4.5 optimizes training and inference for improved text generation, image comprehension, and multi-modal reasoning, catering to diverse real-world applications.

Baidu has released its ERNIE 4.5 open-source large language model series, featuring ten models including Mixture-of-Experts (MoE) variants. These models offer fully open-sourced pre-training weights and inference code, accessible via platforms like Hugging Face. The series boasts innovative heterogeneous MoE architecture for multimodal capabilities and achieves state-of-the-art performance on various benchmarks, outperforming competitors in text and multimodal tasks. The models are distributed under the Apache 2.0 license, promoting both academic and commercial use.

Huawei has open-sourced its Pangu 7B dense and Pangu-Pro MoE 72B large language models, along with Ascend-based inference technology. This move supports Huawei’s Ascend ecosystem strategy, aiming to accelerate AI research and application. The Pangu-Pro MoE 72B model shows strong performance, ranking highly on benchmarks for models under 100 billion parameters.

Moonshot AI launched Kimi-Researcher, its first autonomous AI agent, currently in beta. Built on end-to-end agentic RL, Kimi-Researcher surpasses leading models like Claude 4 Opus and Gemini 2.5 Pro in internal tests, demonstrating strong autonomy and zero-structure adaptability. The agent independently manages research tasks, navigates conflicting information, and prioritizes accurate results. Moonshot AI plans to open-source key components to further accelerate advancements in agentic RL.