AI News

Follow the latest AI and automation news

23 Sept 202612 Rabiʻ II 1448 AH

Inspur MetaBrain SD200 Ultra Packs 128 Domestic AI Chips for 2.8T Kimi K3 Under 5.85ms/Token

Inspur Information introduced the MetaBrain SD200 Ultra AI server at the 2026 Artificial Intelligence Computing Conference, highlighting its capability to handle trillion-parameter workloads. The server integrates 128 domestic AI chips, features 8 TB of unified-address accelerator memory, and 64 TB of host memory, enabling it to run Moonshot AI's Kimi K3 model with 2.8 trillion parameters and a latency under 5.85 milliseconds. Additionally, Inspur claims that AllReduce time improves by about 3.5 times compared to previous methods, significantly enhancing inference performance.

Models

23 Sept 202612 Rabiʻ II 1448 AH

Li Auto Ships Embodied Trilogy: ME-Brain-1.0, ME-U0 Unified WAM and Edge ME-VLM

On September 22, Li Auto's Foundation Model team released an embodied AI trilogy, featuring the ME-Brain-1.0 system, the MachEmbodied-U0 model, and the MachEmbodied-VLM model. This launch emphasizes system and model architecture, focusing on memory, cognition, and action rather than electric vehicle marketing. The ME-Brain-1.0 system integrates Evolvable Memory, a Cognitive Core, and an event-driven Action Model into a loop for execution and skill updates. The Cognitive Core is available in a 35B-A3B version and a 4B variant, with lab tests showing the larger core achieving an average score of 70.9 on embodied suites.

Infrastructure

23 Sept 202612 Rabiʻ II 1448 AH

T-Head Unveils Zhenwu V900 AI Accelerator: About 3× M890, 216GB Memory, Q1 2027 Mass Production

Alibaba's T-Head chip design unit unveiled the Zhenwu V900 at the 2026 Apsara Conference in Hangzhou. This new accelerator boasts approximately three times the performance of the Zhenwu M890, featuring 216 GB of memory and 1,200 GB/s of inter-chip bandwidth. The chip supports multiple data precisions, making it suitable for both high-precision training and low-precision inference. Its larger memory aims to reduce data movement overhead, facilitating the operation of over 1,000 cards in a unified manner.

💡

Airbnb widens access to GPT-6 Astra and OpenAI frontier models

How can AI enhance engineering team efficiency in companies?

Models

OpenAI wants to consult elite mathematicians about how to not fumble again

23 Sept 202612 Rabiʻ II 1448 AH

Models

23 Sept 202612 Rabiʻ II 1448 AH

OpenAI wants to consult elite mathematicians about how to not fumble again

OpenAI announced the formation of an independent panel of mathematicians to assist in improving its interactions with mathematical research. This move follows a reputational crisis due to controversial mathematical results, prompting the company to seek expert guidance. The panel aims to advise on how new results are presented and how to engage with the mathematics community, catching many researchers by surprise. While some view this as a positive first step, there are still fundamental questions about the panel's actual functions.

Business

Grab and OpenAI bring practical AI skills to Southeast Asia

23 Sept 202612 Rabiʻ II 1448 AH

Business

23 Sept 202612 Rabiʻ II 1448 AH

Grab and OpenAI bring practical AI skills to Southeast Asia

OpenAI and Grab have announced the launch of the GO Forward program, aimed at helping 30,000 partners enhance their AI skills across Southeast Asia. The program focuses on providing practical training and support to partners, enhancing their ability to leverage AI technologies in their businesses. This initiative is a significant step towards fostering digital innovation in the region. This initiative comes at a time when reliance on AI is increasing across various sectors, allowing companies to improve their efficiency and competitiveness in the global market.

Business

Snorkel AI triples valuation to $3.5B as demand for AI training data booms

22 Sept 202611 Rabiʻ II 1448 AH

Business

22 Sept 202611 Rabiʻ II 1448 AH

Snorkel AI triples valuation to $3.5B as demand for AI training data booms

The seven-year-old startup has successfully raised $350 million in a Series E funding round. This capital will be used to enhance its data-as-a-service approach. The funding marks a significant step towards expanding the company's service offerings, reflecting the growing demand for data solutions in the market. The startup aims to leverage this investment to advance its technology and improve its services.

Models

Better prompt caching for GPT-6

22 Sept 202611 Rabiʻ II 1448 AH

Models

22 Sept 202611 Rabiʻ II 1448 AH

Better prompt caching for GPT-6

OpenAI has announced new enhancements in the GPT-6 model, significantly improving caching efficiency with higher hit rates. The updates include advanced diagnostic tools and explicit breakpoints, allowing users better control over operations. These features aim to reduce latency and associated costs. These updates represent a crucial step towards enhancing AI model performance, providing users with a smoother and more effective experience. They also empower businesses to utilize these technologies more efficiently.

Models

OpenAI's GPT-6 Sol and Luna cut prices in half but barely move the needle on performance

22 Sept 202611 Rabiʻ II 1448 AH

Models

22 Sept 202611 Rabiʻ II 1448 AH

OpenAI's GPT-6 Sol and Luna cut prices in half but barely move the needle on performance

OpenAI has announced the launch of its GPT-6 Sol and Luna models, which offer similar performance to their predecessors at a lower cost. These new models are priced at half the cost of previous versions, positioning them as direct competitors to Anthropic's pricier offerings. However, independent analyses indicate little actual improvement in performance. With this move, OpenAI aims to strengthen its market share, particularly as Anthropic launched its Opus 5.5 model simultaneously, intensifying competition in the market.

Models

Qualcomm launches two new smartphone chips with emphasis on AI

22 Sept 202611 Rabiʻ II 1448 AH

Models

22 Sept 202611 Rabiʻ II 1448 AH

Qualcomm launches two new smartphone chips with emphasis on AI

Qualcomm announced its new chip capable of running a 30 billion mixture-of-experts model locally. This chip represents an advanced step in artificial intelligence, enabling faster and more efficient data processing. Furthermore, this model could enhance the performance of AI-dependent applications. With this chip, Qualcomm aims to strengthen its position in the AI market, potentially impacting competition with other companies in the field.