Home Blog Page 474

Harbour BioMed Appoints Dr. Jenny Xie as Chief Scientific Officer, Immunology and Head of Global External Innovation

CAMBRIDGE, Mass., ROTTERDAM, Netherlands and SHANGHAI, March 19, 2026 /PRNewswire/ — Harbour BioMed (the “Company”; HKEX: 02142), a global biopharmaceutical company focused on the discovery and development of novel antibody therapeutics in immunology and oncology, today announced the appointment of Dr. Jenny Xie as Chief Scientific Officer, Immunology and Head of Global External Innovation. Dr. Xie will be based in Shanghai and report directly to Dr. Jingsong Wang, Founder, Chairman and Chief Executive Officer of Harbour BioMed.

In this role, Dr. Xie will lead Harbour BioMed’s global external innovation strategy in immunology, managing ongoing strategic collaborations with key partners, as well as identifying and pursuing new partnership opportunities. Additionally, she will oversee all activities at the company’s Harbour BioMed-AstraZeneca Innovation Lab in Beijing.

Dr. Xie brings more than 25 years of experience in the biopharmaceutical industry. Prior to joining Harbour BioMed, she served as Executive Director and Head of Discovery Immunology at Bristol Myers Squibb, where she led research efforts in autoimmune diseases and contributed to the discovery, evaluation, and clinical translation of novel therapeutic approaches. Prior to that, she also held scientific leadership roles at Merck, building extensive experience in early-stage research and development.

Dr. Jingsong Wang, Founder, Chairman, and CEO of Harbour BioMed, commented: “We are very pleased to welcome Dr. Jenny Xie to our executive team. Her extensive experience in drug discovery and development in immunology will be invaluable as we work to strengthen our R&D capabilities and drive innovation across our pipeline. I am confident her leadership will significantly advance our immunology research and global collaboration efforts.”

Dr. Jenny Xie, Chief Scientific Officer, Immunology and Head of Global External Innovation of Harbour BioMed, added: “I am truly excited to join an incredible company like Harbour BioMed and look forward to working with everyone toward our shared goal of transforming patients’ lives through innovation, speed, and global collaboration.”

Dr. Xie received her Ph.D. in Immunology from Boston University School of Medicine and her bachelor’s degree in Biology from Fudan University.

About Harbour BioMed

Harbour BioMed (HKEX: 02142) is a global biopharmaceutical company committed to the discovery and development of novel antibody therapeutics in immunology, oncology and other areas. The Company is building a robust portfolio and differentiated pipeline through internal R&D capability, strategic global collaborations in co-discovery and co-development, and selective acquisitions.

Our proprietary antibody technology platform, Harbour Mice®, generates fully human monoclonal antibodies in both the conventional two heavy and two light chain (H2L2) format and the heavy chain-only (HCAb) format. Building upon HCAb antibodies, the HCAb-based immune cell engagers (HBICE®) bispecific antibody technology enables tumor-killing effects that traditional combination therapies cannot achieve. The HCAb-based Antibody Plus technology (HCAb PLUSTM) provides comprehensive modality solutions for the development of innovative multi-specific medicines in different disease areas. Additionally, building upon the Harbour Mice® platform, Harbour BioMed launched its first fully human Generative AI HCAb Model powered by its Hu-mAtrIxTM AI platform, accelerating the development of innovative therapies.

By integrating Harbour Mice®, HBICE®, HCAb PLUSTM, a single B-cell cloning platform and AI technologies, Harbour BioMed has built a highly efficient and distinctive antibody discovery engine for developing next-generation therapeutic antibodies. For more information, please visit www.harbourbiomed.com.

TEASK, OneQode and Tigasfera Greentech Announce Framework to Deploy Energy and Compute as a Single System

KUALA LUMPUR, Malaysia, March 19, 2026 /PRNewswire/ — As demand for AI and high-performance compute accelerates, three companies are building a model that co-locates energy generation, waste conversion and compute as a single deployable system.

TEASK, OneQode and Tigasfera Greentech today announced a collaboration that combines renewable energy, waste-to-energy systems and high-performance compute into a single deployable model – generating power where infrastructure is needed, converting local waste into additional energy, and running compute directly on top of it.

TEASK provides rapidly deployable solar-powered microgrids. Tigasfera Greentech delivers modular, decentralised waste-to-energy systems that convert biomass waste into renewable electricity and carbon-negative outputs. OneQode delivers sovereign AI cloud compute, telecoms and modular datacenter infrastructure designed to run on these energy sources.

The framework is open and collaborative by design. No single company can solve the converging demands of energy, data and sustainability alone. This approach treats the problem as a system – generation, intelligence and regeneration operating as coordinated layers rather than separate industries.

Malaysia is the collaboration’s initial focus. The country offers strong policy support for renewable energy, a rapidly expanding digital economy, and growing demand for local compute capacity. Tigasfera Greentech, initially incubated through PETRONAS Innovation Garage (PING), is already operating and developing multiple waste-to-energy and carbon credits deployments across Malaysia. TEASK is a Malaysia-based clean energy company building portable, solar-powered energy nodes and EV charging infrastructure. Its systems are designed for rapid deployment in both urban and underserved environments. TEASK is a national award-winning energy company, recognised at the National Energy Awards for renewable energy innovation.

James Tan, CEO, TEASK: “We can deploy a solar energy node in under 30 minutes. That fundamentally changes where infrastructure can exist. When you combine that with waste-to-energy and compute, you’re no longer waiting for the grid to catch up; you’re building the grid, wherever people need it most.”

Matthew Shearing, CEO, OneQode: “We’re on a mission to bring world-class AI compute to countries around the world in a way that’s sovereign, performance-focused, and locally empowering. This collaboration gives us a way to deploy infrastructure in markets where traditional power supply would add years to the timeline.”

Yusof Faizal Amin, CEO, Tigasfera: “By turning waste into power at the source and pair it with digital compute, it brings energy, economic value and digital opportunity directly to local communities. This creates a new model where digital infrastructure doesn’t just consume resources-it uplifts communities and removes carbon as it scales.”

About the companies

TEASK is a Malaysia-based clean energy company building portable solar microgrids and EV charging infrastructure designed for rapid deployment. The company is a national award-winning energy innovator, recognised at the National Energy Awards, and collaborates internationally with partners including Japan’s Leave a Nest.

OneQode is a global infrastructure-as-a-service provider headquartered in Brisbane, Australia. The company operates a latency-optimised carrier network across 30+ countries with 22 points of presence, delivering GPU compute, bare metal servers, cloud platforms and modular datacenter solutions.

Tigasfera Greentech is a Malaysia-based decarbonisation-as-a-service company delivering modular, onsite solutions that convert waste into clean energy, carbon-negative products and verified carbon credits. Incubated by PETRONAS, it intends to be a key enabler of circular economy solutions and carbon-negative infrastructure in the region.

HD Renewable Energy Secures JPY 5.4 Billion Financing for Hokkaido BESS in Japan’s First Grid-Scale Battery Green Bond

TAIPEI, March 19, 2026 /PRNewswire/ — HD Renewable Energy (TPEX: 6873), a leading international smart energy company, announced that its Helios 50MW grid-scale battery energy storage system (BESS) project in Hokkaido, Japan, has secured project financing of approximately JPY 5.4 billion.

HDRE secures project financing for the Hokkaido Helios project, marking Japan’s first Green Project Bond for grid-scale BESS.
HDRE secures project financing for the Hokkaido Helios project, marking Japan’s first Green Project Bond for grid-scale BESS.

The financing was arranged with the support of Nomura Capital Investment and Nomura Securities, and raised through a green project bond, marking Japan’s first green project bond backed by grid-scale battery storage assets. The transaction highlights growing interest from global capital markets in energy storage infrastructure and market-driven power trading models.

The financing adopts a project finance trust beneficiary right structure (ABL), with the HDRE Group serving as the asset manager. The Helios project operates under a merchant model, generating revenue primarily through electricity trading in Japan’s wholesale power market (JEPX) and balancing market (EPRX).

Unlike renewable projects that rely on subsidies or long-term fixed-price contracts, merchant battery storage projects have revenue streams directly linked to market prices, making them a more challenging asset class for traditional project finance. Through a structured operational and cash-flow management framework, the project successfully secured long-term financing with a maturity of up to 19 years, representing an important milestone for battery storage financing in Japan.

As renewable energy penetration continues to rise in Japan, demand for flexible dispatch and grid balancing resources is increasing. Grid-scale battery storage is emerging as a critical infrastructure component to support grid stability and enable higher renewable integration. However, project finance markets have historically favored assets with predictable cash flows, such as FIT-supported projects. Battery storage projects, whose revenues depend on electricity market conditions, have therefore seen relatively limited long-term financing. The successful issuance of the Helios green project bond demonstrates growing investor confidence in energy storage assets and market-based electricity trading models.

“The successful financing of the Helios project marks an important step in establishing a viable financial framework for battery storage assets in Japan,” said Jason Chou, General Manager of HD Renewable Energy. “It also reflects increasing market confidence in power trading-driven business models. As our storage portfolio continues to expand, we will further integrate asset operations with power trading capabilities to enhance the flexibility and value of energy asset management.”

The Helios BESS project, with a capacity of 50MW, is already participating in Japan’s electricity markets and conducting power trading operations. The project is expected to gradually participate in the balancing market and capacity market to further enhance asset performance.

The project has received BBB investment-grade credit and green bond ratings from Rating and Investment Information, Inc. (R&I) and has also obtained a Second-Party Opinion (SPO) confirming its alignment with Green Bond Principles. Going forward, HDRE will continue to strengthen cooperation with international financial institutions including the Nomura Group, while advancing integrated operations of battery storage and power assets and expanding its smart energy footprint in Japan.

Chinese AI Brand Andun Takes Center Stage in New York’s Times Square

Fusing Eastern Wellness Wisdom with Advanced Algorithms to Usher in a “New Era of Proactive Wellness”

NEW YORK, March 19, 2026 /PRNewswire/ — On March 16, Chinese AI tech brand Andun illuminated New York’s Times Square, marking its official introduction to the global market. By pioneering a technological path that integrates the concepts of traditional Chinese wellness with modern artificial intelligence, Andun continues to explore groundbreaking possibilities in the proactive health and lifestyle sector.

Andun made an appearance at the Beijing Artificial Intelligence and Robotics Exhibition
Andun made an appearance at the Beijing Artificial Intelligence and Robotics Exhibition

During the global brand showcase, Andun Founder Bai Weimin emphasized the company’s dedication to driving the evolution of personal well-being through technological innovation. “We hope to help global users better understand their overall wellness status through smarter algorithms and more comprehensive trend analyses,” Bai Weimin stated.

At the core of Andun’s brand ecosystem is its proprietary “Tianhui AI Pulse Analysis Algorithm”. This innovative system seamlessly combines traditional wellness principles with deep learning models to interpret holistic health signals. Rather than relying solely on traditional hardware measurements, Andun places a stronger emphasis on the role of its central algorithms to observe personal wellness trends and generate highly personalized lifestyle, fitness, and nutritional recommendations.

Expanding its proactive wellness ecosystem, Andun highlighted its flagship innovation: the AI Wellness Robot. This smart interactive companion acts as a tangible manifestation of the brand’s powerful algorithm, bringing personalized Eastern wellness insights directly into users’ daily lives. Recently, the robot garnered widespread attention and generated significant buzz at the Beijing Artificial Intelligence and Robotics Conference, drawing crowds and sparking industry conversations.

Prominent Chinese media such as People’s Daily Online and 36Kr noted that Andun’s integration of traditional wellness wisdom and modern technology showcases a promising new direction for independent innovation by Chinese tech enterprises. As the global proactive well-being market continues to evolve, technological paths based on diverse wellness frameworks are expected to collectively drive the industry forward. Through this launch, Andun aims to provide global users with expanded choices, officially pioneering and defining the next generation of smart wellness.

1exchange and Capitaux Partner to List NVDAt, among the First ISIN-Backed Tokenized NVIDIA Stock

SINGAPORE – Media OutReach Newswire – 19 March 2026 – 1exchange, a leading regulated RWA exchange for listing and trading RWA tokens, has entered a strategic partnership with Capitaux, the issuer of NVDAt, an ISIN-registered tokenized security referencing NVIDIA shares. The collaboration is set to bring NVDAt, ISIN-backed tokenized NVIDIA stock, to global investors, establishing a new benchmark for credibility and standardization in the global RWA tokenization market.
NVDAt is structured as a 1:1 asset-backed tokenized representation of NVIDIA shares, designed for compliant and efficient on-chain trading upon listing. ISIN registration enhances transparency, interoperability, and institutional recognizability across trading, settlement, and reporting systems. By supporting issuers in obtaining ISIN registration, 1exchange enables clearer identification and stronger alignment with established financial standards for RWA tokens.
With transparent and compliant architecture, independent third-party custody, and institutional grade blockchain security, Capitaux enables institutions to transform real-world assets into globally tradable digital securities. 1exchange is actively building its listing-to-trading ecosystem to bring high-quality and compliant RWA tokens to investors worldwide. The forthcoming listing of NVDAt further strengthens 1exchange’s tokenized product offering, supporting the scalable growth of the RWA tokenization market.
“By introducing an ISIN-backed RWA token in partnership with Capitaux, we will be bringing institutional-grade compliance standards into the on-chain RWA tokenization market.” says Sheena Lim, CEO of 1exchange, “Our focus is to make the listing and trading of RWA tokens more transparent, structured, and accessible for investors.”

Hashtag: #1exchange #tokenization #toenizedstocks #RWA


The issuer is solely responsible for the content of this announcement.

About 1exchange

1exchange, a member of FOMO Group, is a leading exchange for Real-World Assets (RWA) security tokens and private listings, licensed by the Monetary Authority of Singapore (MAS). Offering full-stack on-chain infrastructure, the platform enables issuers to list enterprise-grade RWAs, while enabling investors to trade modern digital assets in a regulated secondary market, unlocking global liquidity.
Visit for more information. For media inquiries, please contact .


Disclaimer:
The information contained in this article is provided strictly for general informational purposes only. It does not constitute financial advice, investment advice, an offer to sell, or a solicitation of an offer to purchase or subscribe for any securities or financial products listed or traded on 1exchange (“1X”).
Investments involve risks, including the possible loss of principal. Past performance is not necessarily indicative of future performance.
Readers should carefully consider their investment objectives, financial circumstances, and risk tolerance, and should conduct their own independent research. Where appropriate, readers are encouraged to seek advice from a qualified financial professional before making any investment decisions.
This article has not been reviewed by the Monetary Authority of Singapore.

UOB issues RMB 5 billion 3-year Panda bond at one of the lowest coupon for an international issuer

SINGAPORE, March 19, 2026 /PRNewswire/ — UOB has successfully priced an RMB 5.0 billion three-year bond at 1.83 per cent.

UOB was the first Singaporean issuer and only Singaporean bank in the Panda bond market. This is the Bank’s third Panda bond since its debut in 2019 and the sixth Chinese interbank bond offering by the Group, including issuances from UOB China, demonstrating the Group’s long-standing commitment to the capital markets in China. Our Panda bond offering in 2024 was the first time a bank listed a Panda bond on the Singapore Exchange (SGX). Similar to 2024, UOB will also apply for this bond to be listed on SGX subject to regulatory approval.

This offering garnered strong demand with a subscription ratio of 1.6x times from both onshore and offshore investors – 22 per cent of participation from offshore investors, with onshore investors taking 78 per cent. 

This is the joint largest issue size for a three-year tenor among foreign bank issuers, equal to that attained by UOB in 2024. The size achieved is 1bn to 3bn larger than the equivalent 3Y tranches by other foreign bank peers earlier this year.

The final coupon of 1.83 per cent represents a spread of 22 basis points over the three-year China Development Bank benchmark bond yield. The coupon and spread achieved is one of the lowest achieved by an international foreign bank (excl HK) Panda bond issuer to date.

With USD markets largely closed due to geopolitical volatility, the onshore RMB market has remained stable, with onshore issuances remaining normal and yields remaining compressed.

Ms Koh Chin Chin, Head of Group Treasury, Research and Customer Advocacy, UOB, said, “We’re pleased to complete our CNY5 billion Panda bond issuance with strong investor support. As a repeat issuer, we appreciate the depth and resilience of the onshore RMB market, which has remained stable despite a challenging geopolitical backdrop. We remain committed to deepening our long–standing engagement with China’s RMB investor community and supporting the continued growth of the Panda bond market.”

Transaction highlights

  • The trade was marketed with a range of 1.70 to 2.10 per cent saw strong momentum with orderbooks reaching almost 6bn within 3 hours and marginal rate decreasing from to 2.10 per cent to 1.84 per cent by 5pm. There was a 1.1x cover ratio at the final coupon of 1.83%.
  • Final orderbooks reached RMB 8.2 billion with a cover ratio at 1.6x

Distribution

Geography:

  • Onshore investors in China – 78%
  • Offshore investors – 22%

Investor type:

  • Banks – 95%
  • Securities firms – 5%

Bank of China is the Lead Underwriter and Lead Bookrunner, while Industrial and Commercial Bank of China Limited is the Lead Underwriter and Joint Bookrunner, and Agricultural Bank of China Limited, China Construction Bank Corporation, China Galaxy Securities Co., Ltd., China Securities Co., Ltd., CITIC Securities Company Limited, Deutsche Bank (China) Co., Ltd., The Export-Import Bank of China, Fubon Bank (China) Co., Ltd, Industrial Bank Co., Ltd, MUFG Bank (China), Ltd., Postal Savings Bank of China Co., Ltd., Standard Chartered Bank (China) Limited and Sumitomo Mitsui Banking Corporation (China) Limited are Joint Lead Underwriters and Joint Bookrunners.

Laos Phases Out Family Books in Shift to Digital Population System

Laos is replacing Family Books with a digital population management system

Laos is phasing out traditional Family Books as the government accelerates its shift to digital governance.

Under the new policy, authorities will stop issuing hard-copy household registration books. Instead, citizens will use a Certificate of Household Information, generated through a digital system, as the primary document for administrative procedures.

The change is to make it easier to store, access, and manage population data while reducing paperwork and delays in public services, according to a government notice.

Transition Period

Family Books issued in the past remain valid until their expiration date. But people whose household information has already been entered into the digital system must use the new certificate instead when completing official procedures.

Citizens with Digital ID Cards can now use them as official proof of identity for legal transactions and government services, reducing the need for multiple documents.

Only authorised offices can issue the new certificates, including public security departments at central, provincial, and district levels. Authorities are calling on all sectors to recognise the new system and help make the transition smooth.

Building on Digital ID Rollout

The reform builds on Laos’s rollout of digital national ID cards, which expanded nationwide on 17 December 2025 after an initial pilot in Vientiane Capital.

The cards include biometric data and security features such as chips and QR codes, designed to make identity verification faster and more reliable. Officials say the system will support easier access to public services and help connect different government databases.

Together, the changes represent a broader push to move Laos toward fully digital administration, where personal data can be managed more efficiently nationwide.

Breaking the 100M Token Limit: EverMind’s MSA Architecture Achieves Efficient End-to-End Long-Term Memory for LLMs

The research introduces a novel memory architecture called MSA (Memory Sparse Attention). Through a combination of the Memory Sparse Attention mechanism, Document-wise RoPE for extreme context extrapolation, KV Cache Compression with Memory Parallelism, and a Memory Interleave mechanism supporting complex reasoning, MSA achieves a 100-million-token long-term memory framework for LLMs. It delivers industry-leading results on mainstream long-context QA and Needle-In-A-Haystack (NIAH) benchmarks. Remarkably, when scaling the context length from 16K to 100M tokens, the model’s performance degrades by less than 9%, demonstrating extraordinary scalability.

SAN MATEO, Calif., March 19, 2026 /PRNewswire/ — On March 18, EverMind, a pioneer in AI memory infrastructure, released a landmark research paper, Memory Sparse Attention for Efficient End-to-End Memory Model Scaling to 100M Tokens, introducing a novel architecture that enables large language models to achieve efficient, end-to-end long-term memory at the unprecedented scale of 100 million tokens. The paper is published on Zenodo (https://zenodo.org/records/19103670) and open-sourced on GitHub (https://github.com/EverMind-AI/MSA).

Memory Sparse Attention for Efficient End-to-End Memory Model Scaling to 100M Tokens
Memory Sparse Attention for Efficient End-to-End Memory Model Scaling to 100M Tokens

This approach can be viewed as a memory plug-in for large models, providing a fresh perspective and direction for solving the long-term memory problem. In today’s era of exploding Agent ecosystems, this work stands as a potential milestone in ushering in the new epoch of “Memory-as-a-Service”.

MSA integrates topkselection with sparse attention, achieving strong scalability while remaining differentiable
MSA integrates topkselection with sparse attention, achieving strong scalability while remaining differentiable

1.  The “Impossible Triangle” of LLM Long-Term Memory

In recent years, the capability boundaries of LLMs have continuously expanded. However, when it comes to lifelong, fine-grained memory retention, they still face an insurmountable chasm. Scenarios such as literary analysis requiring the comprehension of extensive novel series, Digital Twins demanding persona consistency across multi-turn dialogues, or Agent systems needing to trace long historical records, all place stringent demands on the model’s effective context length. Yet, mainstream LLMs, constrained by the quadratic complexity of Full Attention mechanisms, have long had their effective context windows limited to around 1 million (1M) tokens — a far cry from the estimated hundreds of millions of tokens comprising a human’s lifelong memory capacity.

Memory Sparse Attention layer
Memory Sparse Attention layer

To break through this bottleneck, academia and industry have explored three main technical paradigms. However, while attempting to solve the problem, each paradigm has fallen into new dilemmas, forming an irreconcilable “Impossible Triangle”:

  • Parameter-Based Memory: This method “burns” knowledge directly into model parameters via continuous training or fine-tuning (e.g., LoRA). While it offers high precision, it suffers from poor scalability, high updating costs, and a high susceptibility to catastrophic forgetting.
  • External Storage-Based Memory: Represented by Retrieval-Augmented Generation (RAG), this approach externalizes memory into vector databases. It boasts excellent scalability, but its non-end-to-end, decoupled “retrieve-then-generate” nature makes retrieval precision a performance bottleneck, struggling to achieve deep semantic alignment.
  • Latent State-Based Memory: This paradigm utilizes the model’s internal hidden states (such as KV cache) as working memory. It provides high semantic fidelity but faces a direct conflict between efficiency and capacity. Methods retaining local KV caches offer high precision but limited scalability; whereas linear attention-based methods (like RWKV or DeltaNet) achieve linear complexity but suffer severe precision degradation in ultra-long contexts due to lossy compression.

The following table summarizes the key trade-offs across all three paradigms:

Comparison of Long-Term Memory Methods for LLMs
Comparison of Long-Term Memory Methods for LLMs

Against this backdrop, the MSA paper sets an ambitious goal: to design an end-to-end trainable latent state memory framework that scales to 100M tokens with linear complexity while maintaining high precision. The emergence of MSA aims to directly challenge and shatter the aforementioned “Impossible Triangle,” endowing LLMs with true “lifelong memory.”

2.  Deep Dive into the MSA Architecture: Four Pillars of Innovation

The revolutionary nature of MSA does not stem from a single technological breakthrough, but rather from a cohesive, systemic stack of architectural innovations. These components work in synergy to form the bedrock of its high performance.

2.1  The Core Foundation: Memory Sparse Attention

The core idea of MSA is to introduce a differentiable, content-based sparsification mechanism into the Transformer’s attention layer. Instead of forcing the model to attend to all historical memories during inference, it designs an efficient “Routing” module to dynamically select the most relevant memory subsets for computation.

The MSA layer is wrapped within a standard Pre-Norm Transformer block, replacing the traditional Self-Attention layer. This plug-and-play design ensures seamless integration into existing LLM architectures without disruptive overhauls. The heart of the innovation is an elegant “Dual-Routing” mechanism that selects the best evidence from massive documents without brute-force full computation:

  • Input & Projection: The model receives the external knowledge base and the current query, projecting them into four independent representations: Q (Query), K (Key), V (Value), and a specialized Routing Key (K^R).
  • First-Level Routing (Topic-level Screening): The model performs coarse-grained topic screening across the entire knowledge base using an attention mechanism and token-wise mean pooling. An auxiliary contrastive loss (L_aux) is computed here to enforce routing sparsity and prevent query collapse onto a few popular documents.
  • Second-Level Routing (Token-level Fine-screening): Within the relevant documents identified in the first step, the model conducts finer “content screening” via Pooling and Top-k operations, selecting the most critical sentences or paragraphs at the token level.
  • Final Attention Computation: Only the selected, most essential K and V vectors are loaded into the attention computation. These sparsified memory K/Vs are concatenated with the query’s K/Vs and fed into the final Multi-head Attention layer to generate the output.

The ingenuity of this mechanism lies in internalizing the “retrieval” step of RAG systems into an end-to-end trainable neural network module. Unlike RAG, which relies on external, fixed similarity metrics (like cosine distance), MSA’s router is co-optimized with the generation task during training via a supervised contrastive loss. This fundamentally solves the core pain point of misaligned objectives between RAG’s “retrieval” and “generation,” serving as the key to achieving high precision.

2.2  The Key to Scalability: Document-wise RoPE

To successfully extrapolate from shorter training texts (e.g., 64k tokens) to 100M-level inference texts, handling positional information is a critical challenge. If traditional global positional encoding is used, positional indices will shift drastically when the number of documents during inference far exceeds that during training, leading to severe performance degradation.

MSA proposes a concise and efficient solution: assigning an independent set of Rotary Positional Embeddings (RoPE) to each individual document. This means that regardless of how many documents are in the memory bank, the model’s internal “coordinate system” when reading each document always starts from 0 and remains stable. This design decouples the internal relative position of a document from its absolute position in the global memory, allowing the position-awareness learned during training to generalize losslessly to inference scenarios with massive document banks. Complementing this, Global RoPE is applied to the active query context, with position IDs offset by the number of retrieved documents, ensuring the model perceives the query as a logical continuation of the retrieved background information.

2.3  Engineering Implementation: KV Cache Compression & Memory Parallel

Theoretical feasibility must be translated into reality through engineering. At the 100Mtoken scale, even after compression, the KV cache storage requirement reaches approximately 169 GB — far exceeding the aggregate 160 GB VRAM capacity of a standard 2×A800 node. MSA resolves this physical bottleneck through an ingenious “Memory Parallel” strategy built on two key observations:

  • Tiered Storage: During the routing phase, the model only needs the relatively small Routing Keys (K^R) for quick matching; the massive Content KVs (K and V) are only needed after Top-k selection. Thus, K^R is distributed across GPU VRAM for low-latency retrieval, while K and V are offloaded to CPU DRAM.
  • Asynchronous Fetching: Once the GPU completes routing and determines the Top-k documents, the system asynchronously fetches only the required Content KVs from CPU memory to the GPU for the final generation computation. This “fast search (GPU), slow fetch (CPU)” strategy elegantly shifts the storage bottleneck from limited VRAM to massive CPU memory, enabling 100Mtoken inference on just two A800 GPUs.

2.4  Complex Reasoning Capability: Memory Interleave

For complex questions that require integrating multiple pieces of evidence scattered across different documents (i.e., multi-hop reasoning), a single “retrieve-generate” cycle often falls short. To address this, MSA introduces the Memory Interleave mechanism, which allows the model to perform multiple rounds of “generative retrieval → context expansion” loops:

Three-Stage Inference Process with Memory Interleave
Three-Stage Inference Process with Memory Interleave

In the first round, the model generates the IDs of documents it deems most relevant based on the original query. The system fetches the original text of these documents and appends them to the query, forming a richer “intermediate query.” In the next round, the model generates new document IDs based on this enriched context. This cycle continues until the model determines that the accumulated evidence is sufficient, at which point it switches to generating the final answer.

This iterative reasoning chain simulates the thought process of a human detective: “Discover clue A → Follow the vine to find clue B → Integrate A and B to form a complete evidence chain.” It endows MSA with the ability to dynamically plan its information-gathering path, which is a crucial reason for its outstanding performance on Multi-hop QA tasks. The model autonomously determines how many retrieval rounds are needed per query, rather than relying on a pre-defined fixed number of retrieved documents.

3.  Re-interpreting Experimental Data: Validating MSA’s Value

The paper validates the effectiveness of the MSA architecture from multiple dimensions through exhaustive experiments. We highlight the three most revealing findings.

3.1  Astonishing Scalability and Robustness

On the RULER Needle-In-A-Haystack (NIAH) benchmark, as the context scales from 32K to 1M tokens, MSA’s accuracy only drops from 98.77% to 94.84% — a mere 3.93 percentage-point decline across a 32-fold expansion. In stark contrast, the unmodified Qwen3-4B backbone plummets to 24.69% at 1M tokens, and even the 80B Qwen3-Next model degrades to 80.78% at 1M tokens. In the even more extreme MS MARCO QA test, when expanding the memory scale from 16K to 100M tokens (spanning 4 orders of magnitude), MSA’s performance score only drops from 4.023 to 3.669, a degradation rate of less than 9%. This intuitively proves the architecture’s exceptional robustness against massive irrelevant information (noise) interference.

Results on the "Needle In A Haystack" (NIAH) evaluation across varying context lengths from 32k to 1M tokens.
Results on the “Needle In A Haystack” (NIAH) evaluation across varying context lengths from 32k to 1M tokens.

3.2  The Power of End-to-End Optimization

On the average scores of 9 QA benchmarks, the 4B-parameter MSA model (average score 3.760) significantly outperforms complex RAG systems built on the identical Qwen3-4B foundation, including those with a Reranker. Specifically, MSA achieves average improvements of 16.0%, 11.5%, and 14.8% over standard RAG, RAG with reranking, and HippoRAG2, respectively. More strikingly, on several datasets, its performance even surpasses top-tier RAG systems composed of the SOTA KaLMv2 retriever paired with the massive 235B-parameter Qwen3 model — a system with 58× more parameters. This fully demonstrates the high-precision advantage brought by MSA’s end-to-end optimization.

MSA integrates topkselection with sparse attention, achieving strong scalability while remaining differentiable
MSA integrates topkselection with sparse attention, achieving strong scalability while remaining differentiable

3.3  The Indispensability of Each Component

Ablation studies clearly quantify the contribution of each innovation. Compared to the baseline MSA-S1 model, removing the Memory Interleave mechanism leads to a 5.3% average performance drop, with HotpotQA suffering a 19.2% decline — confirming its essentiality for multi-hop reasoning. Removing the auxiliary routing supervision in Continual Pre-training causes a severe 31.3% average performance degradation (43.1% on HotpotQA), as errors in initial document retrieval compound during subsequent interleaving steps. Removing the Original Text Injection causes the most severe slide of 37.1%, with DuReader experiencing a 46.2% drop, indicating that precise final answers still rely critically on the semantic details of the original text. This proves that MSA is an intricately designed, organic whole where every component is indispensable.

4.  Conclusion: MSA’s Originality and Core Value

In summary, the true value of MSA lies not merely in releasing a powerful long-context model, but in providing the AI memory field with a brand new, fully validated technical infrastructure that simultaneously achieves scalability, precision, and efficiency. It proves that we do not have to make painful compromises between the “low precision” of RAG and the “high cost” of full attention. By cleverly combining the idea of sparsification with the end-to-end learning capabilities of neural networks, building an independent, scalable “Memory Layer” compatible with LLMs is entirely feasible.

This paints an exciting blueprint for the future development of the AI ecosystem: Memory can act as an independent, pluggable service, freely combined with various reasoning cores (LLMs). User data and “memory assets” will no longer be locked into any single model or vendor. From this perspective, MSA is not just an excellent academic paper; it is likely a milestone that inaugurates the new era of Memory-as-a-Service.

5.  Background: EverMind and Shanda Group’s “Discoverative AI” Vision

To fully understand the driving force behind the MSA research, it is necessary to examine it within the macro-strategic context of its creator, EverMind, and its parent company, Shanda Group. EverMind is one of the core teams deeply incubated by Shanda Group’s founder, Tianqiao Chen, in the AI field. Its mission is to conquer the long-term memory challenge of AI, moving towards AI’s Self-Evolving capability.

According to recent interviews with Tianqiao Chen by Bloomberg and TMTPost, Shanda Group’s AI strategy does not focus on the current mainstream “Generative AI,” but aims to build a more pioneering “Discoverative AI” ecosystem. Its ultimate goal is to have AI assist humans in discovering new knowledge and solving fundamental problems like disease and energy, rather than merely imitating and recombining existing information. In this grand vision, two foundational technologies are placed at the core:

  • MiroMind — Reasoning: This team is dedicated to enabling models to proactively seek evidence from the external world and revise hypotheses like scientists, achieving true reasoning reliability and insightful discovery through paths like verifiable reasoning.
  • EverMind — Memory: This team’s mission is to build an infinitely scalable, high-fidelity long-term memory system for AI that is independent of any specific model. Only when AI possesses a stable and reliable memory foundation can it conduct effective, cross-temporal complex reasoning and knowledge creation, ultimately achieving Self-Evolving AI.

Therefore, EverMind and MiroMind jointly constitute the core driving force of Shanda Group’s “Discoverative AI” blueprint, corresponding to the two core pillars of cognitive science: Memory and Reasoning. The MSA architecture introduced in this paper is exactly the core technological achievement of the EverMind team in practicing the “Memory-as-a-Service” concept. Its underlying design and technical route not only represent a breakthrough in the existing long-context bottleneck but also profoundly reflect Shanda Group’s long-term investment and firm determination in building independent, autonomous, and controllable AI infrastructure.