MoE Models

Twitter

Facebook

Copy Link

Key Facts

Year1991, 2017

Key ModelsGPT-5, DeepSeek V3, Llama 4, Mixtral, Qwen3

Key PeopleRobert Jacobs, Michael Jordan (1991); Noam Shazeer (2017)

Top News

Tencent launches Hy3 AI Model with 40% efficiency gain and real-world product integration

Tencent unveiled Hy3 preview, a 295-billion-parameter Chinese LLM that improves inference efficiency by 40% while excelling at coding and agent workflows. Unlike typical model launches focused on benchmarks, Tencent is proving Hy3's capabilities through live integration across Yuanbao, CodeBuddy, and other products, where it reduced latency by 54% and handles complex tasks up to 495 steps.

Technology2 sources

Tencent launches Hy3 AI Model with 40% efficiency gain and real-world product integration

Technology2 sources

Google releases Gemma 4 with Apache 2.0 license, enabling unrestricted local AI on devices

Google has launched Gemma 4, its latest generation of open-weight AI models, marking a significant shift to the Apache 2.0 license from its previous restrictive terms. The release includes four model variants optimized for everything from smartphones to enterprise servers, with the 31B model ranking third on the Arena AI leaderboard. This licensing change removes commercial deployment barriers and positions Gemma 4 as a domestic alternative to Chinese open-weight models.

Technology19 sources

Google releases Gemma 4 with Apache 2.0 license, enabling unrestricted local AI on devices

Technology19 sources

Microsoft pushes to develop state-of-the-art AI models by 2027 after ending OpenAI restrictions

Microsoft AI chief Mustafa Suleyman announced plans to develop cutting-edge AI models by 2027, marking a major shift toward AI self-sufficiency. The move follows a renegotiated OpenAI deal that previously barred Microsoft from building broadly capable models. However, computing power constraints mean the company still can't compete with frontier AI leaders like Anthropic and OpenAI.

Technology6 sources

Microsoft pushes to develop state-of-the-art AI models by 2027 after ending OpenAI restrictions

Technology6 sources

Nvidia launches Nemotron 3 Super with 5x throughput boost to power enterprise AI agents

Nvidia unveiled Nemotron 3 Super, a 120-billion-parameter open model designed to run complex agentic AI systems at scale. The model combines Mamba, Transformer attention, and Latent MoE architectures to deliver up to 5x higher throughput and 2x better accuracy than its predecessor. It addresses context explosion and thinking tax challenges that plague multi-agent workflows, which generate up to 15x more tokens than standard chat interactions.

Technology9 sources

Nvidia launches Nemotron 3 Super with 5x throughput boost to power enterprise AI agents

Technology9 sources

NPCI partners with Nvidia to build sovereign AI infrastructure for India's digital payments

NPCI has partnered with Nvidia to develop a payments-native AI foundation model aligned with India's regulatory and data sovereignty requirements. The collaboration will use Nvidia Nemotron to create a foundational AI layer for the payments ecosystem, building on the UPI Help Assistant pilot that supports grievance resolution at scale.

Technology2 sources

NPCI partners with Nvidia to build sovereign AI infrastructure for India's digital payments

Technology2 sources

NVIDIA Blackwell Ultra slashes AI inference costs by 35x while delivering 50x better performance

NVIDIA's GB300 NVL72 systems powered by Blackwell Ultra GPUs achieve up to 50x higher throughput per megawatt and 35x lower cost per token compared to the Hopper platform. Cloud providers including Microsoft, CoreWeave, and Oracle are deploying these systems at scale for agentic AI and coding assistants, while leading inference providers report 4x to 10x cost reductions using open-source models.

Technology6 sources

NVIDIA Blackwell Ultra slashes AI inference costs by 35x while delivering 50x better performance

Technology6 sources

Microsoft pushes for AI self-sufficiency as it builds in-house models to reduce OpenAI dependence

Microsoft's AI chief Mustafa Suleyman announced the company is pursuing AI self-sufficiency by developing its own advanced foundation models. The strategic shift aims to reduce reliance on OpenAI while maintaining their partnership through 2032. Microsoft previewed MAI-1, trained on 15,000 NVIDIA H100 GPUs, and launched the Maia 200 inference chip to control AI economics.

Business3 sources

Microsoft pushes for AI self-sufficiency as it builds in-house models to reduce OpenAI dependence

Business3 sources

AI in Chemistry breakthrough: Two systems accelerate molecule design and drug discovery by 10x

Researchers unveiled two AI systems that transform chemical synthesis and molecule design. Yale's MOSAIC successfully helped synthesize 35 new compounds for pharmaceuticals and materials, while NYU's PropMolFlow generates molecular candidates ten times faster than previous tools. Both systems address critical bottlenecks in drug discovery and could accelerate the path from laboratory concept to real-world products.

Science2 sources

AI in Chemistry breakthrough: Two systems accelerate molecule design and drug discovery by 10x

Science2 sources

Moonshot AI releases Kimi K2.5 with Agent Swarm that coordinates 100 parallel AI agents

Chinese AI startup Moonshot AI unveiled Kimi K2.5, an open-source large language model that generates web interfaces from visual inputs and coordinates up to 100 sub-agents working in parallel. The model outperforms proprietary models from OpenAI, Anthropic, and Google on key benchmarks while being trained on 15 trillion tokens.

Technology5 sources

Moonshot AI releases Kimi K2.5 with Agent Swarm that coordinates 100 parallel AI agents

Technology5 sources

Phison aiDAPTIV+ delivers 10X faster AI inference on consumer PCs using NAND flash as memory

Phison showcased aiDAPTIV+ at CES 2026, a software and hardware solution that accelerates AI inference by up to 10 times on consumer PCs. The technology uses NAND flash as memory expansion, enabling systems with just 32GB of DRAM to run AI models requiring 96GB through conventional approaches. Partners including Nvidia, AMD, MSI, and Acer demonstrated the technology, which addresses the memory bottleneck limiting AI deployment on ordinary hardware.

Technology3 sources

Phison aiDAPTIV+ delivers 10X faster AI inference on consumer PCs using NAND flash as memory

Technology3 sources

DeepSeek's Engram tackles AI compute waste with conditional memory breakthrough

DeepSeek unveiled Engram, a conditional memory technique that separates static information retrieval from complex reasoning tasks. The method allows AI models to bypass GPU memory constraints by committing knowledge to system RAM, reducing reliance on expensive high-bandwidth memory while improving performance on long-context queries.

Technology5 sources

DeepSeek's Engram tackles AI compute waste with conditional memory breakthrough

Technology5 sources

Nvidia launches Nemotron 3 open source AI models as Meta steps back from transparency

Nvidia released its Nemotron 3 family of models in three sizes—Nano (30B), Super (100B), and Ultra (500B parameters)—with full training data transparency. The move positions the chip giant as a major force in open source AI just as Meta signals a retreat from openness, leaving Chinese models to dominate the space.

Technology16 sources

Nvidia launches Nemotron 3 open source AI models as Meta steps back from transparency

Technology16 sources

NVIDIA AI Servers Achieve 10x Performance Boost for Mixture of Experts Models from DeepSeek and Moonshot

NVIDIA has shattered performance records with its GB200 Blackwell NVL72 servers, delivering a tenfold efficiency boost for mixture-of-expert AI models including those from China's DeepSeek and Moonshot AI. The breakthrough addresses a critical computing bottleneck in scaling MoE models, using 72 chips with high-bandwidth links and 30TB of fast shared memory to enable expert parallelism at unprecedented levels.

Technology3 sources

NVIDIA AI Servers Achieve 10x Performance Boost for Mixture of Experts Models from DeepSeek and Moonshot

Technology3 sources

Mistral AI launches 10 open-source models to challenge OpenAI and Google with edge computing focus

French AI startup Mistral AI released its Mistral 3 family featuring 10 open-source AI models, including a 675B-parameter frontier model and nine smaller variants designed for edge devices. The release positions smaller, customizable models as superior alternatives to large closed-source systems, with capabilities spanning robotics, drones, and offline environments while operating on single GPUs.

Technology11 sources

Mistral AI launches 10 open-source models to challenge OpenAI and Google with edge computing focus

Technology11 sources

Baidu Unveils Next-Generation AI Chips to Challenge Western Dominance in Chinese Market

Chinese search giant Baidu announces new AI training and inference accelerators, including the M100 and M300 chips, alongside rack-scale systems designed to reduce reliance on Western semiconductors amid escalating US-China tech tensions.

Technology2 sources

Baidu Unveils Next-Generation AI Chips to Challenge Western Dominance in Chinese Market

Technology2 sources

Prev 123 Next