LLM Daily: May 29, 2025
🔍 LLM DAILY
Your Daily Briefing on Large Language Models
May 29, 2025
HIGHLIGHTS
• Nvidia reported 69% revenue growth in Q1 despite incurring a $4.5 billion charge due to licensing requirements for its H20 AI chip in China, highlighting the impact of geopolitical tensions on AI chip markets.
• DeepSeek has quietly released their latest MIT-licensed model, DeepSeek-R1-0528, which is receiving high praise from users and is already being converted to GGUF format by Unsloth for easier local deployment.
• Researchers from Tel Aviv University and Meta AI have developed PISCES, a groundbreaking method that can selectively erase specific concepts from LLMs without degrading general performance, addressing privacy, copyright, and safety concerns.
• Open-source LLM development platforms are seeing significant growth, with Dify reaching nearly 100K GitHub stars and Langflow gaining 600+ stars in a single day, demonstrating strong community interest in accessible AI application frameworks.
BUSINESS
Nvidia Reports Strong Q1 with 69% Revenue Growth Despite China Challenges
Nvidia continues to dominate the AI chip market with impressive financial results, though regulatory challenges are affecting its China business. CEO Jensen Huang expressed concerns about U.S. policy restricting AI chip sales to China, as the company incurred a $4.5 billion charge in Q1 due to licensing requirements for its H20 AI chip. Despite these challenges, Nvidia's overall financial performance remains strong. TechCrunch, 2025-05-28
Mistral AI Launches New Products to Compete with Industry Leaders
Mistral AI has made two significant product announcements:
-
Codestral Embed: A new code embedding model that reportedly outperforms offerings from OpenAI and Cohere in real-world retrieval tasks. The model aims to improve RAG use cases and help identify duplicate code segments using natural language. VentureBeat, 2025-05-28
-
Mistral Agents API: A comprehensive API for building AI agents capable of running Python, generating images, performing RAG, and more. This positions Mistral as a competitor in the growing AI agent infrastructure market. VentureBeat, 2025-05-27
Chinese AI Lab DeepSeek Gains Market Attention
DeepSeek has rapidly gained prominence with its chatbot app climbing to the top of both the Apple App Store and Google Play charts. The company has also updated its R1 reasoning AI model and released it on Hugging Face under an MIT license, allowing for commercial use. Industry analysts are now questioning whether the U.S. can maintain its competitive edge in AI development. TechCrunch, 2025-05-28
Anthropic Expands Claude's Capabilities with Voice Mode
Anthropic has begun rolling out a voice mode for its Claude chatbot mobile apps, allowing users to have spoken conversations with the AI assistant. The feature will also integrate with Google Docs, Drive, and Calendar, broadening Claude's functionality and accessibility. The voice mode is currently in beta and will be available in English over the coming weeks. TechCrunch, 2025-05-27
OpenAI Exploring "Sign in with ChatGPT" for Third-Party Apps
OpenAI is developing a service that would allow users to sign in to third-party applications using their ChatGPT accounts. The company is currently gauging interest from developers who might want to integrate this authentication service into their apps, potentially leveraging ChatGPT's growing user base to expand its ecosystem. TechCrunch, 2025-05-27
PRODUCTS
DeepSeek Releases R1-0528: New MIT-Licensed Model Without Fanfare
DeepSeek AI has quietly released their latest open-source language model, DeepSeek-R1-0528 (2025-05-28), maintaining their MIT license commitment. The new model has garnered significant attention on Reddit, with users praising its capabilities as "lethal" and commenting on DeepSeek's low-key approach to releasing advanced models without excessive marketing. The model is already being converted to GGUF format by the Unsloth team for easier local deployment.
Unsloth Working on DeepSeek-R1-0528 GGUF Conversion
Unsloth (2025-05-28) is actively converting DeepSeek's latest model to GGUF format, making it more accessible for local deployment and use on consumer hardware. This third-party support highlights the vibrant ecosystem around open-source AI models and the rapid community response to new releases.
Llama 4 Identified as Underserved by Inference Providers
Despite being a major open-source model release, Llama 4 has been identified as underserved by inference providers and API platforms according to discussion on r/MachineLearning. This suggests an opportunity for services to improve accessibility and deployment options for Meta's latest foundation model, which could help expand its adoption and use cases.
TECHNOLOGY
Open Source Projects
Dify - LLM App Development Platform
Dify is a comprehensive open-source platform for building LLM applications with nearly 100K GitHub stars. It offers an intuitive interface that combines AI workflow management, RAG pipelines, agent capabilities, and model management with robust observability features. Recent updates include removing agent turn limits and fixing token length issues for LLM invocations.
Langflow - Visual Agent Builder
Langflow (65K+ stars) provides a powerful visual tool for building and deploying AI-powered agents and workflows. Recent commits show active development with the addition of a models category and improvements to the chat widget. The project has seen significant growth with nearly 600 new stars in a single day.
Lobe Chat - Multi-Provider AI Chat Framework
Lobe Chat is an open-source chat framework (61K+ stars) featuring modern design and support for multiple AI providers including OpenAI, Claude 4, Gemini, Ollama, and DeepSeek. It supports knowledge bases with file upload functionality, multi-modal interactions, and can be deployed with a single click. Recent fixes include resolving OIDC authorization issues on Windows 11.
Models & Datasets
DeepSeek-R1-0528
DeepSeek's latest model release has quickly gained 626 likes on Hugging Face. The model focuses on conversational abilities and includes custom code generation capabilities, with optimizations for FP8 precision deployment.
BAGEL-7B-MoT
ByteDance's BAGEL-7B-MoT model has attracted over 800 likes and 5,800+ downloads. Built on Qwen2.5-7B-Instruct, this "any-to-any" model implements the Mixture of Thoughts approach described in their recent arxiv paper (2505.14683).
Devstral-Small-2505
Mistral AI's Devstral-Small-2505 has become an instant hit with 629 likes and 120K+ downloads. This multilingual model supports 25+ languages including English, French, German, Spanish, Japanese, Korean, Chinese, and many more. It's optimized for vLLM deployment with Apache 2.0 licensing.
Mixture-of-Thoughts Dataset
This training dataset (111 likes, 5,600+ downloads) implements the Mixture of Thoughts approach for enhancing LLM reasoning. The collection contains between 100K-1M examples in parquet format and is associated with research papers (2504.21318, 2505.00949).
EuroSpeech Dataset
EuroSpeech is a multilingual speech dataset with 75 likes and an impressive 33K+ downloads. It covers 24 European languages and supports both automatic speech recognition and text-to-speech tasks, containing between 1-10M examples in parquet format.
Developer Tools & Interfaces
Kolors-Virtual-Try-On
This Gradio-based application by Kwai has amassed nearly 9,000 likes on Hugging Face Spaces. It enables virtual clothing try-on, allowing users to visualize how different garments would look without physically wearing them.
AI Comic Factory
With over 10,000 likes, this Docker-based space lets users create AI-generated comics. The application streamlines the comic creation process by automating illustration and layout based on user inputs.
Chatterbox
Resemble AI's Chatterbox (69 likes) is a Gradio-based conversational interface for voice interaction. It leverages Resemble's voice synthesis technology to create natural-sounding AI conversations.
Step1X-3D
This Gradio space from stepfun-ai has accumulated 206 likes and showcases 3D generation capabilities. The tool demonstrates the latest advancements in AI-powered 3D model creation from text descriptions or reference images.
Background Removal
A practical utility with 1,900 likes, this Gradio space offers efficient background removal from images. It provides a simple interface for extracting subjects from their backgrounds without requiring advanced editing skills.
RESEARCH
Paper of the Day
Precise In-Parameter Concept Erasure in Large Language Models (2025-05-28)
Authors: Yoav Gur-Arieh, Clara Suslik, Yihuai Hong, Fazl Barez, Mor Geva
Institution: Tel Aviv University, Meta AI Research
This paper stands out for introducing PISCES, a groundbreaking method for selectively erasing specific concepts from LLMs' knowledge without degrading general performance. Unlike previous approaches that use fine-tuning or adapter methods, PISCES precisely identifies and modifies only the parameters that encode the targeted concept, making it a significant advancement for addressing privacy, copyright, and safety concerns in deployed models.
The researchers demonstrate PISCES can successfully erase specific knowledge (like medical information about certain medications) while maintaining model performance on unrelated tasks. Their evaluation shows the method works effectively across multiple model architectures and sizes, representing an important step toward more controlled and responsible AI deployment.
Notable Research
Topological Structure Learning Should Be A Research Priority for LLM-Based Multi-Agent Systems (2025-05-28)
Authors: Jiaxi Yang, Mengqi Zhang, Yiqiao Jin, et al.
This position paper makes a compelling case for shifting research focus toward optimizing the network topology of LLM-based multi-agent systems, arguing that how agents are structurally organized has critical implications for collaborative performance but remains underexplored compared to agent capability development.
Zero-Shot Vision Encoder Grafting via LLM Surrogates (2025-05-28)
Authors: Kaiyu Yue, Vasu Singla, Menglin Jia, et al.
The researchers introduce a novel approach for efficiently training vision-language models by using smaller "surrogate" language models during vision encoder training, then transferring the encoder to larger LLMs without additional training, reducing computational costs while maintaining performance.
Agent-UniRAG: A Trainable Open-Source LLM Agent Framework for Unified Retrieval-Augmented Generation Systems (2025-05-28)
Authors: Hoang Pham, Khac-Hoai Nam Bui
This paper presents a unified framework for RAG systems using LLM agents as controllers, addressing both single-hop and multi-hop reasoning queries through a trainable architecture that shows improved performance over traditional RAG approaches in complex question-answering tasks.
DocReRank: Single-Page Hard Negative Query Generation for Training Multi-Modal RAG Rerankers (2025-05-28)
Authors: Navve Wasserman, Oliver Heinimann, Yuval Golbari, et al.
The authors introduce an innovative approach for training multimodal RAG rerankers by actively generating hard negative examples from single documents rather than relying on passive retrieval, creating more diverse and challenging training data that improves reranker performance across various benchmarks.
LOOKING AHEAD
As we approach Q3 2025, the convergence of multimodal models with specialized reasoning frameworks is poised to redefine AI capabilities. The recent breakthroughs in neural-symbolic integration demonstrated by DeepMind and Anthropic suggest we'll see models with significantly enhanced logical reasoning abilities by year-end, potentially transforming scientific research and complex decision-making processes.
Meanwhile, the regulatory landscape continues to evolve rapidly. With the EU's AI Act implementation deadlines looming in Q4 and similar frameworks emerging in Asia, we anticipate a standardization push toward explainable AI architectures. Companies that have invested in transparency tools and responsible deployment practices will likely gain competitive advantages as these regulations solidify across global markets.