🔍 LLM DAILY

Your Daily Briefing on Large Language Models

November 30, 2025

HIGHLIGHTS

• Supabase has achieved a remarkable $5 billion valuation after raising $100 million, more than doubling its valuation in just months as it becomes the backend of choice for AI applications and developer tools.

• NeKot has emerged as a significant new terminal-based interface that supports both local and cloud LLMs, offering comprehensive features including major AI service integration and image support without requiring additional runtime.

• Google's open-source Gemini CLI has garnered over 85,000 stars on GitHub, bringing Google's Gemini directly to developers' terminals with recent improvements to hook systems and telemetry.

• University of Tübingen researchers have published the first comprehensive evaluation of model merging techniques for LLMs, revealing that subspace-based merging methods outperform traditional weight-averaging approaches for combining model capabilities.

BUSINESS

Supabase Reaches $5B Valuation with New Funding Round

Supabase, the open-source database platform that has become the backend of choice for vibe-coding applications, raised $100 million at a $5 billion valuation. This comes just months after securing $200 million at a $2 billion valuation, showing remarkable growth momentum. The company has positioned itself as critical infrastructure in the AI and developer tools ecosystem. TechCrunch (2025-11-28)

Michael Burry Takes Position Against Nvidia

Michael Burry, known for predicting the 2008 housing market collapse, has positioned himself against Nvidia, creating significant market discussion. According to TechCrunch, questions remain about whether Burry is correctly forecasting an inevitable collapse in Nvidia's valuation or if his growing influence might actually trigger the market movement he's predicting. Nvidia has been one of the primary beneficiaries of the AI boom through its GPU dominance. TechCrunch (2025-11-27)

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

Representatives from OpenAI and Google shared insights at TechCrunch Disrupt about how AI is transforming go-to-market strategies for technology companies. The discussion focused on how investors and startups are adapting their product introduction approaches in response to rapid advancements in artificial intelligence capabilities. TechCrunch (2025-11-28)

Online Black Friday Spending Reaches Record $11.8 Billion

According to Adobe Analytics, which tracks more than 1 trillion visits to U.S. retail websites, American consumers spent a record $11.8 billion online during Black Friday. The data suggests continued growth in e-commerce, increasingly powered by AI-driven personalization and recommendation systems. TechCrunch (2025-11-29)

PRODUCTS

NeKot: A Terminal Interface for Local and Cloud LLMs

GitHub Repository | Developer: Balance Software | Released: 2025-11-29

A new terminal-based interface for interacting with both local and cloud-based large language models has gained significant attention in the local LLM community. NeKot, developed by an independent developer, offers a comprehensive solution for users seeking a lightweight but feature-rich way to interact with AI models. Key features include:

• Support for major AI services including Gemini, OpenAI, and OpenRouter APIs
• Compatibility with popular local LLM solutions (llama-cpp with llamaswap, Ollama, LM Studio)
• Image support capabilities
• Customizable presets with dedicated settings and system prompts
• Session management
• Basic vim motion support for efficient navigation

Written in Go, the application requires no additional interpreter or runtime, making it particularly attractive for users seeking an efficient, resource-light solution for LLM interactions. The project appears to address a gap in the market for maintained terminal interfaces with comprehensive feature sets for both cloud and local AI model access.

The developer noted they created the tool after being unable to find an existing solution that wasn't abandoned and had all their required features, suggesting a potential need in the developer community for more sustainable LLM interaction tools.

TECHNOLOGY

Open Source Projects

google-gemini/gemini-cli

An open-source AI agent that brings Google's Gemini directly to your terminal. With over 85,000 stars, this TypeScript-based tool enables developers to interact with Gemini models through a command-line interface, streamlining AI assistance during development workflows. Recent updates focus on comprehensive hook system testing and improved telemetry for tracking response events.

firecrawl/firecrawl

A powerful Web Data API for AI applications that converts websites into LLM-ready markdown or structured data. This TypeScript project (68,805 stars) serves as a crucial bridge between web content and LLMs, making it easier to process and analyze web information. Recent commits show active development, including switching to UUIDv7 and improving error handling with Sentry integration.

pathwaycom/llm-app

Ready-to-run cloud templates for RAG systems, AI pipelines, and enterprise search with live data synchronization. This Docker-friendly project (47,668 stars) provides seamless integration with various data sources including Sharepoint, Google Drive, S3, Kafka, and PostgreSQL. The repository is being actively reorganized to improve template accessibility and documentation.

Models & Datasets

Tongyi-MAI/Z-Image-Turbo

A high-performance text-to-image diffusion model with 1,322 likes and over 31,000 downloads. Z-Image-Turbo implements the custom ZImagePipeline in the Diffusers framework and has been published with research backing (arxiv:2511.13649), offering advanced image generation capabilities under an Apache 2.0 license.

black-forest-labs/FLUX.2-dev

A versatile image generation and editing model with 720 likes and over 168,000 downloads. FLUX.2-dev features a unique Flux2Pipeline in Diffusers for both image-to-image and generation workflows, providing developers with flexible image manipulation capabilities.

tencent/HunyuanOCR

Tencent's multimodal OCR model with 512 likes and nearly 60,000 downloads. This transformers-based model supports image-text-to-text conversion and is compatible with both English and Chinese text recognition. The model is documented in arxiv:2511.19575 and is compatible with AutoTrain and HuggingFace Endpoints.

nvidia/PhysicalAI-Autonomous-Vehicles

NVIDIA's autonomous vehicle dataset with 420 likes and over 148,000 downloads. This recently updated dataset (November 29) provides essential training data for developing and testing autonomous driving systems, supporting the growing field of physical AI applications.

opendatalab/AICC

A large multilingual text dataset (between 1-10B samples) designed for text generation tasks. With 45 likes and over 13,000 downloads, this CC-BY-4.0 licensed dataset is compatible with multiple libraries including datasets, dask, mlcroissant, and polars. The dataset focuses on web corpus content and is documented in arxiv:2511.16397.

Developer Tools & Infrastructure

HuggingFaceTB/smol-training-playbook

A comprehensive Docker-based guide for small model training with 2,474 likes. This research-focused space provides visualization tools and templates for scientific papers, helping developers efficiently train smaller, more resource-friendly AI models with best practices documentation.

burtenshaw/karpathy-llm-council

A Gradio-based implementation of Andrej Karpathy's LLM Council concept with 80 likes. This space leverages multiple language models to create a "council" that can deliberate and reach consensus on complex tasks, providing an innovative approach to ensemble learning and decision-making with LLMs.

facebook/sam3

Meta's latest Segment Anything Model with 779 likes and over 234,000 downloads. SAM3 extends the capabilities of previous segmentation models to video content, providing transformers-based feature extraction and mask generation. The model is compatible with HuggingFace Endpoints, making deployment more accessible.

deepseek-ai/DeepSeek-Math-V2

A specialized mathematical reasoning model with 463 likes and over 2,600 downloads. This Apache 2.0 licensed model builds on DeepSeek's previous work, with optimizations for conversational mathematical problem-solving. The model supports various deployment options including AutoTrain and Endpoints compatibility, with FP8 quantization support.

RESEARCH

Paper of the Day

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

Oğuz Kağan Hitit, Leander Girrbach, Zeynep Akata

University of Tübingen

This paper stands out for providing the first comprehensive evaluation of model merging techniques specifically for LLMs, filling a critical knowledge gap in the field. The authors conduct a large-scale systematic evaluation of six state-of-the-art merging methods across four open-weight LLMs and twelve fine-tuned checkpoints per base model, offering valuable insights for practitioners.

The research reveals that while model merging can effectively combine capabilities of different fine-tuned models, the benefits vary significantly based on the specific merging method, task combination, and base model. Notably, the study finds that certain subspace-based merging techniques outperform traditional weight-averaging approaches, and that merging can sometimes yield models that exceed the performance of their constituent parts. These findings have important implications for efficient model reuse and deployment in resource-constrained environments.

Notable Research

ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration (2025-11-26) Hongjin Su, Shizhe Diao, Ximing Lu, et al. (NVIDIA, UW, Allen Institute for AI) ToolOrchestra introduces a novel framework that dynamically orchestrates multiple specialist models and tools, achieving 70-80% performance of GPT-4 while using only 1/50th of the parameters, demonstrating a path toward more efficient and practical AI systems.

MADRA: Multi-Agent Debate for Risk-Aware Embodied Planning (2025-11-26) Junjian Wang, Lidan Zhao, Xi Sheryl Zhang This paper presents a training-free framework that leverages collective reasoning through multi-agent debate to assess risks in embodied AI planning tasks, achieving a 20% improvement in hazard detection compared to single-agent approaches while maintaining task completion rates.

Revisiting Generalization Across Difficulty Levels: It's Not So Easy (2025-11-26) Yeganeh Kordi, Nihal V. Nayak, Max Zuo, Ilana Nguyen, Stephen H. Bach This research challenges existing assumptions about LLM generalization across task difficulties, demonstrating that models typically generalize better to examples of similar difficulty levels as their training data, with important implications for data curation strategies.

Tool-RoCo: An Agent-as-Tool Self-organization Large Language Model Benchmark in Multi-robot Cooperation (2025-11-26) Ke Zhang, Xiaoning Zhao, Ce Zheng, et al. The authors introduce a novel benchmark that evaluates LLMs' ability to engage in long-term multi-agent cooperation by treating other agents as tools, providing a framework to assess autonomous agent cooperation without predefined orchestration.

LOOKING AHEAD

As 2025 draws to a close, we're witnessing the maturation of multimodal reasoning systems that seamlessly integrate visual, auditory, and linguistic inputs at human-expert levels. The Q1 2026 release calendar is packed with domain-specific AI systems showing unprecedented capabilities in scientific discovery, particularly in materials science and drug development, where AI-first research teams are outpacing traditional approaches by orders of magnitude.

The regulatory landscape is poised for significant shifts in early 2026, with the EU's AI Harmonization Framework taking effect and similar coordinated policies expected from the U.S.-APAC AI Alliance. Watch for emerging tensions between open and closed AI development paradigms as the performance gap narrows and enterprise adoption of fully-autonomous decision systems accelerates across critical infrastructure sectors.