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

Google's DiffusionGemma AI model generates text 4x faster with parallel processing approach

Google DeepMind released DiffusionGemma, an experimental open-source AI model that generates text in parallel blocks rather than sequentially. The model achieves speeds exceeding 1,000 tokens per second on NVIDIA H100 GPUs and 700+ tokens per second on RTX 5090 hardware. While it trades some output quality for speed, it excels at structured tasks like code infilling and constraint-heavy problems.

Technology10 sources

Google's DiffusionGemma AI model generates text 4x faster with parallel processing approach

Technology10 sources

Phison aiDAPTIV and Intel tackle local AI's memory problem with 26B-parameter models on 16 GB RAM

Phison unveiled aiDAPTIV at Computex 2026, a memory extension technology that runs 26-billion-parameter models on PCs with just 16 GB RAM. Developed with Intel, the system uses high-performance NAND flash storage as an AI cache to bypass memory and storage limitations. But the specialized Pascari SSDs cost $2,516 for 1TB, raising questions about adoption and echoing Intel's past struggles with proprietary technologies like Optane.

Technology2 sources

Phison aiDAPTIV and Intel tackle local AI's memory problem with 26B-parameter models on 16 GB RAM

Technology2 sources

NVIDIA launches Agent Toolkit with NemoClaw to build secure autonomous AI agents at scale

NVIDIA unveiled its Agent Toolkit at GTC Taipei, featuring NemoClaw blueprints and OpenShell runtime for building secure, long-running AI agents. Major companies like Cadence, Siemens, and Synopsys are using the framework to create autonomous AI engineers that compress weeks of simulation work into hours across semiconductor design, industrial engineering, and manufacturing workflows.

Technology6 sources

NVIDIA launches Agent Toolkit with NemoClaw to build secure autonomous AI agents at scale

Technology6 sources

NVIDIA Cosmos 3 Brings Multimodal AI to Robots, Autonomous Vehicles and Industrial Systems

NVIDIA launched Cosmos 3, an open world foundation model for physical AI that combines vision reasoning, multimodal generation and action prediction. Trained on 20 trillion tokens including nearly a billion images and 400 million videos, the model helps robots and autonomous vehicles understand causal relationships and predict outcomes before acting in real-world environments.

Technology6 sources

NVIDIA Cosmos 3 Brings Multimodal AI to Robots, Autonomous Vehicles and Industrial Systems

Technology6 sources

Thinking Machines unveils interaction models that respond in 0.40 seconds for real-time AI chat

Thinking Machines, the AI startup founded by former OpenAI CTO Mira Murati, announced interaction models that process input and generate responses simultaneously. The TML-Interaction-Small model responds in 0.40 seconds—matching natural human conversation speed—and uses full-duplex architecture to listen, see, and talk in real time. A limited research preview arrives in coming months.

Technology6 sources

Thinking Machines unveils interaction models that respond in 0.40 seconds for real-time AI chat

Technology6 sources

Astera Labs unveils Scorpio X switch to tackle AI compute cluster networking bottlenecks

Astera Labs introduced Scorpio X, a vendor-agnostic AI fabric switch featuring 320 lanes of PCIe 6.0 connectivity and 5.12 TB/s bandwidth. The chip aims to solve networking bottlenecks in massive AI compute clusters by enabling GPUs to access distributed resources like local memory. Production is expected in the second half of 2026.

Technology2 sources

Astera Labs unveils Scorpio X switch to tackle AI compute cluster networking bottlenecks

Technology2 sources

Nvidia releases Nemotron 3 Nano Omni with 30B params for edge AI agents and multimodal tasks

Nvidia launched Nemotron 3 Nano Omni, an open-weight multimodal AI model that unifies vision, audio, and language understanding in a single architecture. With 30 billion parameters but only 3 billion active per inference, the model delivers 9x higher throughput than comparable models while running on a single GPU. Early adopters include Foxconn, Palantir, and H Company, with Dell, Oracle, and Infosys evaluating it for production deployment.

Technology6 sources

Nvidia releases Nemotron 3 Nano Omni with 30B params for edge AI agents and multimodal tasks

Technology6 sources

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

Prev12 Next