Top AI Updates of the Week (September 2025, Week 2)

We have rounded up the most important developments in the AI world this week. Here are the most notable AI updates from the second week of September 2025:

[1] Qwen3-ASR — All-in-One Speech Recognition Model

Qwen3-ASR is a next-generation all-in-one automatic speech recognition model developed by Alibaba's Qwen team. The model is seen as a significant step forward in speech recognition.

[2] Google Launches AI Mode in 5 Languages

Google has rolled out its “AI Mode” in search in five new languages in addition to English: Hindi, Indonesian, Japanese, Korean and Brazilian Portuguese. This expansion makes the AI-powered search experience more accessible to users worldwide.

[3] High-Memory A100 Runtimes in Google Colab

Google Colab has introduced “High-Memory A100 Runtimes.” This new feature lets users run NVIDIA A100 GPUs with significantly increased RAM capacity.

[4] Baidu Releases ERNIE X1.1

Baidu has released ERNIE X1.1, its latest reasoning model. The model shows significant improvements in factuality, instruction following and agentic capabilities.

[5] G42 Releases the K2 Think Reasoning LLM

K2 Think is an advanced reasoning LLM introduced by MBZUAI and G42. Although it is built with only 32 billion parameters, this open-source system delivers impressive performance.

[6] mmBERT - A Multilingual Model Based on ModernBERT

mmBERT is a state-of-the-art multilingual encoder model built on the ModernBERT architecture. It was trained on 3 trillion tokens covering more than 1,800 languages.

[7] Agents Can Now Learn with the ART Library

The Agent Reinforcement Trainer (ART) library is an open-source framework designed to simplify and accelerate reinforcement learning (RL) training for Large Language Model (LLM) agents.

[8] MCP Server Tools Now in ChatGPT

“MCP Server Tools in ChatGPT” is a powerful integration that lets ChatGPT connect directly to external MCP (Model Context Protocol) servers, specifically in developer mode.

[9] LLM Checkpoint Engine

Introduced by MoonshotAI, “Checkpoint Engine” is a middleware solution designed to dynamically update model weights inside LLM inference engines.

[10] Anthropic Introduces the Web Fetch Tool

The Anthropic Web Fetch Tool lets Claude retrieve and analyze the full content of specified web pages and PDF documents directly through the API.

[11] Meta Introduces MobileLLM-R1

MobileLLM-R1 is a family of next-generation efficient reasoning models designed by Meta AI for mobile and edge devices.

[12] Google Introduces VaultGemma-1B

VaultGemma-1B is the latest release in Google's Gemma family and was trained from scratch with strict differential privacy guarantees.

[13] Google Announces Genkit Go 1.0

Genkit Go 1.0 is Google's open-source AI development framework designed specifically for the Go programming language. It offers a unified interface to multiple AI model providers such as Google AI, OpenAI and Anthropic.

---

This week brought some truly exciting developments in the AI world. Progress in multilingual support, mobile optimization and open-source solutions stands out in particular. We will keep following how these developments evolve in the coming weeks.

Don't forget to follow our blog so you never miss the latest developments in the AI world!