🔍 LLM DAILY

Your Daily Briefing on Large Language Models

November 29, 2025

HIGHLIGHTS

• Supabase has secured $100M at a $5B valuation, marking explosive growth as the backend infrastructure provider of choice for the emerging vibe-coding segment, highlighting the increasing investor interest in AI infrastructure.

• Unsloth has successfully quantized Alibaba Cloud's powerful Qwen3-Next-80B model down to just 3 billion parameters, making enterprise-grade AI accessible on consumer hardware with reasonable inference speeds.

• Google's open-source Gemini-CLI project has garnered over 85,000 GitHub stars, bringing Gemini's AI capabilities directly to terminal users with recent updates to hook systems and telemetry improvements.

• Researchers from the University of Tübingen have published the first large-scale systematic evaluation of model merging techniques for LLMs, finding performance improvements of up to 13% compared to single models but challenging assumptions about which merging methods work best.

BUSINESS

Supabase Raises $100M at $5B Valuation, Riding Vibe-Coding Wave

TechCrunch (2025-11-28)
Open-source database platform Supabase has secured $100 million at a $5 billion valuation, just months after raising $200 million at a $2 billion valuation. According to TechCrunch, the company has become the backend infrastructure of choice for the booming vibe-coding segment. Supabase's rapid valuation growth highlights the increasing investor interest in AI infrastructure providers supporting new coding paradigms.

OpenAI and Google Share Insights on AI's Impact on Go-to-Market Strategies

TechCrunch (2025-11-28)
At TechCrunch Disrupt, experts from OpenAI and Google discussed how AI is transforming product market strategies for both startups and investors. The companies shared perspectives on how artificial intelligence is reshaping traditional go-to-market approaches across various sectors.

Investor Michael Burry Takes Position Against Nvidia

TechCrunch (2025-11-27)
Famed investor Michael Burry, known for predicting the 2008 financial crisis, has positioned himself against AI chip giant Nvidia. TechCrunch reports that market watchers are debating whether Burry's stance signals an impending correction in AI-related stocks or if his own influence could trigger the market shift he's predicting.

49 US AI Startups Raised $100M+ in 2025

TechCrunch (2025-11-26)
A TechCrunch analysis reveals that 49 AI startups in the United States have each raised at least $100 million in 2025, demonstrating continued strong investment in the sector. The report examines how 2025's funding landscape compares to the previous year's record-breaking investments in artificial intelligence.

xAI Building Solar Farm Near Memphis Data Center

TechCrunch (2025-11-26)
Elon Musk's AI company xAI is developing a solar farm adjacent to its Colossus data center in Memphis. According to TechCrunch, the 88-acre solar installation is expected to generate approximately 30 megawatts of electricity, which would cover about 10% of the data center's estimated power consumption. This move reflects growing efforts by AI companies to address the environmental impact of energy-intensive computing operations.

PRODUCTS

New AI Model Releases

Unsloth Releases Qwen3-Next-80B Quantized to 3B for Local Use (2025-11-28)

Unsloth has released quantized versions of Alibaba Cloud's Qwen3-Next-80B model, compressed to 3 billion parameters for local deployment. The release includes both Instruct and Thinking variants, making this powerful model accessible to users with consumer hardware. Reddit users are testing whether these models can run effectively on setups like a 12GB 4070 Super with 64GB RAM, hoping to achieve reasonable inference speeds.

Z Image Turbo Gets LoRA Training Support (2025-11-28)

Ostris has released a new toolkit enabling users to train LoRA adapters for the Z Image Turbo model. The developer announced this on Twitter/X and has made a training adapter available on Hugging Face. This development significantly expands the model's customization capabilities, with users already reporting successful character LoRA training for Z-Turbo. The community has responded enthusiastically to this advancement, with many expressing excitement about the new creative possibilities.

TECHNOLOGY

Open Source Projects

google-gemini/gemini-cli

An open-source AI agent that brings the power of Gemini directly into your terminal. With over 85,000 GitHub stars, this TypeScript project enables seamless terminal-based interactions with Google's Gemini AI models. Recent updates include comprehensive hook system integration testing and telemetry improvements with OpenTelemetry API response event tracking.

firecrawl/firecrawl

A specialized web data API designed for AI applications that transforms entire websites into LLM-ready markdown or structured data. This TypeScript project has gained significant traction with nearly 69,000 stars. Recent developments include switching to UUIDv7 and adding Sentry integration for better exception monitoring, making it a valuable tool for web data extraction workflows.

pathwaycom/llm-app

Ready-to-run cloud templates for RAG, AI pipelines, and enterprise search with live data synchronization. This Docker-friendly project offers seamless integration with Sharepoint, Google Drive, S3, Kafka, PostgreSQL, and real-time data APIs. Recent updates have focused on restructuring the project organization, moving pipelines to templates and fixing relative links to assets.

Models & Datasets

Tongyi-MAI/Z-Image-Turbo

A high-performing text-to-image diffusion model from Tongyi-MAI with over 1,100 likes and 16,700+ downloads. The model is built on the diffusers framework and uses a custom ZImagePipeline for image generation. Its popularity is also evident in its corresponding Gradio space, which has garnered 550 likes.

black-forest-labs/FLUX.2-dev

A versatile image generation and editing model using the diffusion approach, with significant community adoption (692 likes and 162,000+ downloads). FLUX.2-dev features a specialized Flux2Pipeline for various image manipulation tasks, supporting both image-to-image and direct generation workflows.

tencent/HunyuanOCR

Tencent's multilingual OCR model built on the Hunyuan VL platform. With nearly 500 likes and 41,500+ downloads, this transformer-based model excels at image-text-to-text conversions and is compatible with various deployment platforms including AutoTrain and Hugging Face Endpoints.

nvidia/PhysicalAI-Autonomous-Vehicles

A comprehensive dataset for autonomous vehicle research from NVIDIA, attracting 413 likes and over 142,000 downloads. This dataset provides essential training data for physical AI applications in autonomous driving systems.

nex-agi/agent-sft

A supervised fine-tuning dataset for agent development containing between 10,000 and 100,000 examples in both English and Chinese. With its ODC-BY license, this dataset (62 likes, 573 downloads) provides valuable training data for text-generation tasks focused on agent behaviors.

Developer Tools

HuggingFaceTB/smol-training-playbook

An exceptionally popular developer resource with 2,465 likes, this Docker-based space provides a comprehensive playbook for efficient model training. It combines research article formatting with data visualization capabilities, making it an essential reference for AI practitioners looking to optimize their training workflows.

facebook/sam3

Meta's latest iteration of the Segment Anything Model, SAM3 expands capabilities to include video understanding with mask generation. With 764 likes and over 208,000 downloads, this transformer-based model provides advanced feature extraction capabilities and is compatible with Hugging Face Endpoints for easy deployment.

deepseek-ai/DeepSeek-Math-V2

The second version of DeepSeek's specialized mathematics model, built on transformer architecture with safetensors support. This model (429 likes, 1,019 downloads) is optimized for mathematical problem-solving and features compatibility with AutoTrain, Endpoints, and FP8 precision, making it highly versatile for various deployment scenarios.

Infrastructure

opendatalab/AICC

A massive multilingual web corpus dataset (size between 1-10 billion samples) designed for text generation models. This dataset (39 likes, 7,550 downloads) features extensive Common Crawl data with HTML parsing and conversion to web-corpus and markdown formats. It's accessible through multiple data processing libraries including datasets, dask, mlcroissant, and polars.

burtenshaw/karpathy-llm-council

An implementation of Andrej Karpathy's LLM Council concept with 74 likes, hosted as a Gradio application. This space demonstrates a practical approach to creating consensus-driven AI decision systems by combining outputs from multiple language models, offering insights into ensemble techniques for improved AI responses.

Wan-AI/Wan2.2-Animate

A highly popular animation tool (2,579 likes) implemented as a Gradio application. This space showcases advanced animation capabilities for AI-generated content, providing creators with tools to add motion to static images and create dynamic visual content.

RESEARCH

Paper of the Day

A Systematic Study of Model Merging Techniques in Large Language Models (2025-11-26)

Authors: Oğuz Kağan Hitit, Leander Girrbach, Zeynep Akata

Institution: University of Tübingen

This paper stands out as the first large-scale systematic evaluation of model merging techniques for LLMs, addressing a critical gap in our understanding of how well these methods transfer from smaller models to large language models. By thoroughly testing six state-of-the-art merging approaches across multiple model architectures and fine-tuned checkpoints, the authors provide actionable insights that challenge common assumptions in the field.

The study reveals that while model merging can improve performance by up to 13% compared to single models, the effectiveness varies significantly depending on the base model, dataset properties, and merging method used. Notably, the researchers found that more complex subspace methods don't necessarily outperform simpler weight averaging approaches in the LLM context, contradicting previous findings from smaller models. These results provide important practical guidance for researchers and practitioners seeking to efficiently combine specialized model capabilities without additional training.

Notable Research

ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration (2025-11-26)

Authors: Hongjin Su, Shizhe Diao, Ximing Lu, et al.

ToolOrchestra introduces a novel framework that efficiently combines multiple LLMs and specialized tools, strategically assigning different components of complex tasks to the most suitable resources, demonstrating significant performance improvements while reducing computational costs compared to monolithic approaches.

MADRA: Multi-Agent Debate for Risk-Aware Embodied Planning (2025-11-26)

Authors: Junjian Wang, Lidan Zhao, Xi Sheryl Zhang

This paper presents a training-free framework that employs multiple LLM agents in structured debates to assess risks in embodied planning tasks, significantly improving safety by reducing both false positives and false negatives compared to single-agent approaches while maintaining high task completion rates.

Revisiting Generalization Across Difficulty Levels: It's Not So Easy (2025-11-26)

Authors: Yeganeh Kordi, Nihal V. Nayak, Max Zuo, Ilana Nguyen, Stephen H. Bach

The authors challenge existing assumptions about LLM generalization across task difficulties, presenting a systematic evaluation that reveals models generally perform best when tested on examples matching the difficulty of their training data, with important implications for dataset curation and model evaluation.

Tool-RoCo: An Agent-as-Tool Self-organization Large Language Model Benchmark in Multi-robot Cooperation (2025-11-26)

Authors: Ke Zhang, Xiaoning Zhao, Ce Zheng, et al.

This research introduces a novel benchmark that treats cooperative agents as tools, enabling evaluation of LLMs' capabilities for autonomous multi-robot cooperation without predefined orchestration, providing new insights into how language models handle complex collaborative planning and execution tasks.

LOOKING AHEAD

As we approach 2026, the integration of multimodal reasoning with physical-world interaction capabilities is poised to redefine AI applications. The Q1 2026 releases from major labs are expected to showcase systems that can seamlessly interpret and manipulate real-world environments through advanced robotic interfaces, moving beyond today's primarily cognitive applications.

Meanwhile, the regulatory landscape continues to evolve, with the EU's AI Act Phase 3 implementation and the anticipated U.S. Federal AI Framework both scheduled for early 2026. These developments, coupled with recent breakthroughs in computational efficiency that have reduced training costs by 40%, suggest we're entering an era where specialized, task-optimized AI systems will become more accessible to mid-size enterprises. Watch for this democratization to accelerate innovation across previously underserved sectors in the coming quarters.