ForProduction

// Act Wisely: Cultivating Meta-Cognitive Tool Use in Agentic Multimodal Models // A framework for developing meta-cognitive capabilities in multimodal agents, enabling them to strategically select and use tools based on task requirements. Key highlights: Cultivates meta-cognitive tool selection, improves strategic decision-making in multimodal agents, enables adaptive tool usage based on context, and enhances agent autonomy in complex tasks. By training agents to reflect on their own capabilities and limitations, it enables more intelligent tool selection and reduces reliance on hardcoded workflows or human intervention. 📄 Paper arxiv.org/pdf/2604.08545 💻 Code github.com/Accio-Lab/Metis

// FP4 Explore, BF16 Train: Diffusion Reinforcement Learning Rollout Scaling // A novel approach to diffusion RL that uses FP4 precision for exploration and BF16 for training, addressing computational bottlenecks in rollout scaling. Key highlights: Combines FP4 exploration with BF16 training, reduces computational overhead in diffusion RL rollouts, maintains training stability while accelerating inference, and optimizes the trade-off between precision and performance. By decoupling exploration precision from training precision, it achieves faster rollout generation without sacrificing training quality or model convergence. 📄 Paper arxiv.org/pdf/2604.06916 💻 Project Page nvlabs.github.io/Sana/Sol-RL/

// nezha: Run Multiple AI Coding Agents Across Projects // A framework for running multiple AI coding agents across different projects simultaneously, supporting both Claude Code and Codex. Key highlights: 86 stars, 1 contributor, orchestrates multiple coding agents in parallel, supports Claude Code and Codex, and enables distributed development workflows across multiple codebases. By coordinating multiple AI agents across different projects, it enables parallel development and code review without manual context switching or agent reconfiguration. 🔗 Repo github.com/hanshuaikang/n…

// MiniMax-M2.7: Next-Generation Multimodal Language Model // An advanced multimodal language model from MiniMaxAI featuring improved reasoning capabilities, enhanced visual understanding, and optimized performance for diverse applications. Key highlights: Enhanced multimodal processing, improved reasoning and instruction following, optimized for both chat and complex task completion, and designed for efficient deployment across various use cases. By combining advanced architecture with comprehensive training on diverse multimodal data, it delivers stronger performance on reasoning, coding, and visual tasks while maintaining efficient inference for production workloads. 🤗 Model huggingface.co/MiniMaxAI/Mini…

// MolmoWeb: Open Visual Web Agent and Open Data for the Open Web // An open visual web agent framework and dataset for training web automation models on open web data rather than proprietary sources. Key highlights: Open visual web agent architecture, provides open training data for web automation, focuses on publicly accessible web content, and enables reproducible web agent research without proprietary datasets. By providing open datasets and agent architectures for web automation, it democratizes web agent development and enables researchers to build models that interact with the open web without relying on closed, proprietary data sources. 📄 Paper arxiv.org/pdf/2604.08516 💻 Project Page allenai.org/blog/molmoweb

// auto-deep-researcher-24x7: Autonomous Research Agent // An autonomous AI agent that runs deep research tasks 24/7 while you sleep, featuring a zero-config setup and Leader-Worker architecture. Key highlights: Runs continuously without supervision, Leader-Worker architecture for distributed research tasks, 162 stars, 4 issues, and designed to automate literature review and data collection. By delegating research tasks to autonomous agents that work around the clock, it enables comprehensive literature reviews and data gathering without manual effort or time zone constraints. 🔗 Repo github.com/Xiangyue-Zhang…

@NousResearch @Xiaomi Nothing is free, our data is the price of admission. Hope everyone used the free trail responsibly.

We're glad Hermes users have been making use of the free MiMo V2 Pro access via the Nous Portal! You loved it so much that we faced heavier initial usage than anticipated. Thank you to @Xiaomi for helping us improve the stability - update Hermes and it should now be rock solid.

// SkillClaw: Let Skills Evolve Collectively with Agentic Evolver // A framework that enables skills to evolve collectively through an autonomous agentic evolver, addressing the static nature of skills in existing LLM agent systems. Key highlights: Continuous skill aggregation from multi-user interactions, autonomous skill evolution through agent feedback, prevents skill stagnation after deployment, and enables cross-user skill improvement propagation. By treating skill evolution as a first-class concern and implementing an autonomous evolver that refines skills based on real-world usage patterns, it enables LLM agents to continuously improve and adapt rather than remaining fixed after initial deployment. 📄 Paper arxiv.org/pdf/2604.08377

// claude-obsidian: Persistent Knowledge Companion for Claude + Obsidian // A Claude Code plugin that builds and maintains a persistent, compounding wiki vault in Obsidian. Every source you add gets integrated, every question pulls from everything that has been read. Key highlights: Based on Karpathy's LLM Wiki pattern, supports `/wiki` `/save` `/autoresearch` `/canvas` commands, auto-updates hot cache between sessions, and includes pre-configured Dataview dashboards + CSS snippets. By extracting entities, updating cross-references, and maintaining a structured index, it enables knowledge to compound like interest — answers cite specific wiki pages, not training data, and the vault stays healthy without manual cleanup. 🔗 Repo github.com/AgriciDaniel/c…

@kimmonismus Didn't Iran already declined a few days ago?

edit: still waiting on irans reply

Ceasefire with Iran via cnn

// Memory Intelligence Agent: Enhanced Agent Memory Systems // A research framework for improving agent memory systems through structured memory augmentation and retrieval mechanisms for long-term context retention. Key highlights: Proposes novel memory augmentation techniques, enhances long-term context retention, improves agent decision-making through better memory access, and addresses limitations of standard context windows. By implementing structured memory systems that go beyond simple context windows, it enables agents to maintain coherent long-term reasoning and recall relevant information across extended interactions. 📄 Paper arxiv.org/pdf/2604.04503 💻 Code github.com/ECNU-SII/MIA

// knowledge-engine: Bridge Between Human and Machine Memory // A knowledge engine that bridges human-readable wiki patterns with machine-speed memory retrieval, built on Karp Wiki pattern and Memvid architecture. Key highlights: 26 stars, 2 contributors, combines human-readable documentation with fast machine access, and implements wiki-style knowledge organization with optimized retrieval. By translating between human-readable wiki formats and machine-optimized memory structures, it enables rapid knowledge access while maintaining human-friendly documentation standards. 🔗 Repo github.com/tashisleepy/kn…

// Self-Execution Simulation Improves Coding Models // A training approach that improves coding models by simulating code execution during training, enabling models to better understand program behavior and correctness. Key highlights: Simulates program execution during training, improves competitive programming performance, combines execution feedback with natural language explanations, and addresses the gap between code generation and actual program behavior. By incorporating execution simulation into the training process, it enables models to learn from the actual outcomes of generated code rather than just surface-level patterns, leading to more reliable and correct code generation. 📄 Paper arxiv.org/pdf/2604.03253

// claw-code: Claude Code Snapshot for Research // A snapshot of Claude Code preserved for research purposes, providing access to the original source code for academic study and analysis. Key highlights: 123 stars, 247 forks, research-focused snapshot, preserves original Claude Code implementation, and enables comparative analysis of AI coding assistants. By maintaining an archived version of Claude Code, it enables researchers to study the evolution of AI coding tools and conduct reproducible experiments on code generation models. 🔗 Repo github.com/emmarktech/cla…

⚠️ Better than TurboQuant ⚠️ // RotorQuant: KV Cache Compression via Block-Diagonal Rotation // A drop-in KV cache quantization library that bypasses the butterfly network using block-diagonal rotations, delivering better perplexity, faster decode, and significantly fewer parameters than Google's TurboQuant. Key highlights: 28% faster decode speed, 5.3x faster prefill, 44x fewer parameters (128 vs 16,384), matches or beats TurboQuant on PPL (6.91 vs 7.07 at 3-bit), and includes production-ready llama.cpp integration with PlanarQuant (2D Givens) and IsoQuant (4D quaternion) backends. By replacing dense d×d orthogonal transforms with simple 2D/4D block rotations, it achieves O(d) complexity instead of O(d log d), enabling fully parallelizable quantization that preserves directional structure in KV cache vectors while drastically reducing compute overhead. 🔗 Repo github.com/scrya-com/roto… 📄 Paper scrya.com/rotorquant.pdf 🌐 Site scrya.com/rotorquant

// TriAttention: Efficient Long Reasoning with Trigonometric KV Compression // A novel KV cache compression method that addresses memory bottlenecks in LLMs by leveraging Q/K vector concentration in pre-RoPE space for efficient long-context generation. Key highlights: Exploits Q/K concentration around fixed non-zero centers, uses trigonometric series for distance-based key scoring, achieves 2.5x higher throughput or 10.7x KV memory reduction, and matches Full Attention accuracy on AIME25 with 32K-token generation. By estimating key importance through pre-RoPE centers and trigonometric distance preferences, it enables running 32B models with long context on single consumer GPUs (24GB) where Full Attention would cause out-of-memory errors. 📄 Paper arxiv.org/abs/2604.04921 🔗 Code github.com/nvidia/TriAtte…

// MinerU2.5-Pro: Pushing the Limits of Data-Centric Document Parsing // An advanced document parsing system that pushes the boundaries of data-centric approaches for processing complex documents at scale. Key highlights: Data-centric document parsing, handles complex document structures, processes at scale, and improves upon previous MinerU versions with enhanced accuracy and robustness. By focusing on data-centric methods rather than just model improvements, it achieves superior document parsing performance across diverse document types and formats. 📄 Paper huggingface.co/papers/MinerU2…

// Test-Time Scaling Makes Overtraining Compute-Optimal // A research paper examining how test-time scaling affects model training efficiency and the compute-optimality of overtraining strategies. Key highlights: Analyzes test-time scaling laws, examines overtraining compute-optimality, and provides insights into modern LLM training dynamics and sample efficiency trade-offs. By understanding the relationship between test-time scaling and training compute, it enables more efficient model development strategies that balance performance with resource constraints. 📄 Paper huggingface.co/papers/Test-Ti…

// Deep-Dive-Claude-Code: Production Code Analysis // A comprehensive analysis tool that breaks down Claude Code's production code layer by layer to identify potential issues and architecture problems. Key highlights: Layer-by-layer code analysis, identifies production issues, 80 stars, 34 forks, and designed to detect problems like data leaks and production readiness. By systematically examining each layer of the codebase, it enables developers to catch architectural issues and potential security problems before deployment. 🔗 Repo github.com/waiterxiaoyy/D…