LLM Daily: September 25, 2025
🔍 LLM DAILY
Your Daily Briefing on Large Language Models
September 25, 2025
HIGHLIGHTS
• Cohere has secured $100 million in new funding, reaching a $7 billion valuation, while simultaneously forming a strategic partnership with AMD that strengthens its position in the enterprise AI market.
• Microsoft is diversifying its AI partnerships by integrating Anthropic's technology into its Copilot platform, potentially signaling a gradual shift away from its exclusive reliance on OpenAI.
• Alibaba's QWEN IMAGE GEN and QWEN EDIT 2509 have demonstrated impressive capabilities for generating dynamic widescreen videos from a single static image, showcasing the rapid advancement of AI-powered video creation tools.
• Microsoft's educational initiative "AI-agents-for-beginners" has gained massive community traction with over 39,000 GitHub stars, highlighting the growing interest in accessible AI agent development resources.
• The "Pathways of Thoughts" research from Google introduces multi-directional thinking for LLMs, mimicking human cognition to significantly improve personalized question answering with long, noisy contexts.
BUSINESS
Cohere Reaches $7B Valuation, Partners with AMD
Cohere has secured a fresh $100 million investment (2025-09-24), boosting its valuation to $7 billion just one month after its previous funding round. According to TechCrunch, the enterprise AI company has also announced a strategic partnership with AMD, signaling its continued momentum in the competitive AI market.
Microsoft Integrates Anthropic's AI into Copilot
Microsoft is adding Anthropic's AI technology to its Copilot platform (2025-09-24), TechCrunch reports. This move represents a significant development in Microsoft's AI strategy and potentially marks another step in the gradual separation between Microsoft and OpenAI, as the tech giant diversifies its AI partnerships.
Oracle Reportedly Seeking $15B Through Bond Sale
Oracle is looking to raise approximately $15 billion through a corporate bond sale (2025-09-24), according to TechCrunch. This financial move comes shortly after Oracle reportedly signed a massive $300 billion compute deal with OpenAI, highlighting the company's aggressive investment in AI infrastructure.
OpenAI Expanding Data Center Infrastructure
OpenAI is building five new "Stargate" data centers (2025-09-23) in partnership with Oracle and SoftBank, TechCrunch reports. This substantial infrastructure expansion aims to support the training and deployment of increasingly powerful AI models, reflecting OpenAI's growing computing needs and long-term ambitions.
Neon App Pays Users for Voice Data, Sells to AI Firms
Neon, currently the second most popular app on Apple's App Store (2025-09-24), has gained significant traction with its business model of paying users to record their phone calls and then selling this voice data to AI companies, according to TechCrunch. The app highlights the growing market for high-quality training data as AI companies seek to improve their speech recognition and generation capabilities.
PRODUCTS
QWEN IMAGE GEN & QWEN EDIT 2509: Dynamic Video Generation from Single Images
- Company: Alibaba (established tech company)
- Release Date: (2025-09-24)
- Source: Reddit Post by -Ellary-
Alibaba's QWEN IMAGE GEN and the newly released QWEN EDIT 2509 are being used to create dynamic widescreen videos from a single source image. A Reddit user demonstrated how these tools can be integrated with WAN 2.2 FLF in a Comfy workflow to generate impressive video animations from static images. The demonstration shows how a single image can be transformed into a full video with text effects, transition effects, and maintained text clarity. Community reception has been positive, with users describing the results as "dope" and inquiring about the workflow and required computing specifications.
Raspberry Pi AI Agent: Fully Local LLM System
- Company: Independent developer project (syxa)
- Release Date: (2025-09-24)
- Source: Reddit Post by syxa
An independent developer has created a fully local AI agent system designed specifically for Raspberry Pi 5. This compact system integrates wake-word detection, speech transcription, and LLM inference - all running directly on the Pi hardware without requiring cloud connections. The system utilizes lightweight but capable models like Qwen3:1.7b and Gemma3:1b to function within the Pi's hardware constraints. The project demonstrates the growing feasibility of running complex AI systems on edge devices with limited computing power. Community members have suggested further optimizations including potentially integrating Google's Gemma 3n, which is specifically optimized for CPU-only usage.
Google Gemma 3n: CPU-Optimized Language Model
- Company: Google (established tech company)
- Release Date: (Exact date not specified, mentioned in 2025-09-24 discussion)
- Source: Google AI Dev Documentation
Google has released Gemma 3n, a variant of their Gemma language model series that has been heavily optimized for CPU-only usage. This optimization makes it particularly suitable for deployment on edge devices and environments without GPU acceleration. The model was mentioned in discussions about local AI implementations, particularly in the context of Raspberry Pi applications, highlighting growing interest in efficient AI deployment on resource-constrained devices.
TECHNOLOGY
Open Source Projects
AUTOMATIC1111/stable-diffusion-webui
A comprehensive web interface for Stable Diffusion implemented with Gradio, offering numerous image generation capabilities. With over 156,800 GitHub stars, it features outpainting, inpainting, color sketch, prompt matrix, and upscaling tools in a user-friendly UI. The project remains actively maintained with recent commits focused on bug fixes.
microsoft/ai-agents-for-beginners
A structured educational course from Microsoft containing 12 lessons to help beginners build AI agents. This repository has garnered significant attention with 39,541 stars and nearly 13,000 forks, demonstrating strong community interest in AI agent development fundamentals.
Models & Datasets
ibm-granite/granite-docling-258M
A document understanding model capable of processing complex documents with code, formulas, charts, tables, and layouts. This IDEFICS3-based model excels at document parsing, extraction, and OCR tasks, with over 30,000 downloads and 648 likes. It's specifically designed to handle multi-modal document analysis challenges.
Wan-AI/Wan2.2-Animate-14B
A 14B parameter animation model with 17,800+ downloads that powers the popular Wan2.2-Animate space on Hugging Face (510 likes). The model supports diffusers and ONNX formats, making it versatile for animation generation tasks.
InternRobotics/OmniWorld
A large-scale dataset supporting multiple robotics and multi-modal AI tasks including text-to-video, image-to-video, and image-to-3D conversion. With over 17,600 downloads, this dataset uses the webdataset format and is accompanied by research documented in arXiv:2509.12201.
LucasFang/FLUX-Reason-6M
A reasoning-focused dataset with 41,450 downloads containing 6M+ examples of multi-modal content. Published with an Apache 2.0 license and available in Parquet format, this dataset supports multiple data science libraries including datasets, dask, mlcroissant, and polars.
Developer Tools & Interfaces
not-lain/background-removal
A popular Gradio-based tool for automatic background removal from images with over 2,300 likes. The space provides an accessible interface for a technical image segmentation task that typically requires specialized knowledge.
Kwai-Kolors/Kolors-Virtual-Try-On
An extremely popular virtual try-on application with nearly 9,700 likes, allowing users to visualize clothing items on models. Built with Gradio, this space demonstrates the practical application of image generation for e-commerce and fashion.
yonigozlan/Transformers-Timeline
A visual timeline interface that tracks the evolution of transformer models in AI. With 38 likes, this educational tool helps developers understand the historical development and relationships between different transformer architectures.
RESEARCH
Paper of the Day
Pathways of Thoughts: Multi-Directional Thinking for Long-form Personalized Question Answering (2025-09-23)
Authors: Alireza Salemi, Cheng Li, Mingyang Zhang, Qiaozhu Mei, Zhuowan Li, Spurthi Amba Hombaiah, Weize Kong, Tao Chen, Hamed Zamani, Michael Bendersky
Institution: Google Research, University of Michigan
This paper introduces a groundbreaking approach to personalized question answering by implementing a multi-directional thinking framework that more closely mimics human cognition. The significance lies in its ability to effectively extract and leverage user preferences from long, noisy contexts while generating responses that balance factual accuracy with personalization.
The researchers propose Pathways of Thoughts (PoT), which enables LLMs to explore multiple reasoning directions simultaneously rather than following a single chain of thought. Their experiments show PoT significantly outperforms strong baselines on a new benchmark for personalized long-form QA, demonstrating particular effectiveness in scenarios requiring complex reasoning about user preferences embedded in extensive context.
Notable Research
Extracting Conceptual Spaces from LLMs Using Prototype Embeddings (2025-09-23)
Authors: Nitesh Kumar, Usashi Chatterjee, Steven Schockaert A novel method for deriving interpretable vector spaces from LLMs that align with human cognitive frameworks. The approach uses prototype embeddings to extract conceptual spaces with dimensions that correspond to meaningful perceptual features, offering a promising foundation for more explainable AI.
OnePiece: Bringing Context Engineering and Reasoning to Industrial Cascade Ranking System (2025-09-22)
Authors: Sunhao Dai, Jiakai Tang, et al. This paper addresses the challenge of applying LLM innovations to recommendation systems by combining context engineering and reasoning with traditional ranking architectures. OnePiece demonstrates significant improvements over strong Deep Learning Recommendation Models in industrial settings at Tencent.
Investigating Traffic Accident Detection Using Multimodal Large Language Models (2025-09-23)
Authors: Ilhan Skender, Kailin Tong, Selim Solmaz, Daniel Watzenig A comprehensive study exploring the zero-shot capabilities of MLLMs for traffic accident detection from visual data. The researchers evaluate various leading multimodal models, revealing their strengths and limitations for safety-critical applications in transportation infrastructure.
Data Efficient Adaptation in Large Language Models via Continuous Low-Rank Fine-Tuning (2025-09-23)
Authors: Xiao Han, Zimo Zhao, Wanyu Wang, et al. The authors introduce a novel fine-tuning method that continuously updates a low-rank adaptation matrix during inference, allowing LLMs to efficiently adapt to new tasks and domains with minimal data. This approach significantly improves performance while reducing computational costs compared to traditional fine-tuning.
LOOKING AHEAD
As we close Q3 2025, the AI landscape continues its rapid evolution. We're seeing early applications of quantum-enhanced LLMs from major labs, with IBM and Google leading the race to deploy the first commercial models with significant quantum advantages by early 2026. Meanwhile, the emergence of fully autonomous AI research agents—capable of designing, running, and interpreting their own experiments—suggests we may reach a critical inflection point in AI self-improvement cycles.
Looking toward Q4 and beyond, we anticipate regulatory frameworks will struggle to keep pace with these developments. The EU's AI Sovereignty Act, expected in November, will likely address concerns around autonomous agent deployment. Industry experts are also closely watching advancements in neuromorphic computing architectures, which promise dramatic efficiency improvements that could make today's trillion-parameter models seem quaint by this time next year.