MarkTechPost
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Meta introduced Sapiens2 — a unified computer vision model for pose, segmentation, and 3D
Meta Reality Labs released Sapiens2 — a family of vision models for human analysis that solves pose, segmentation, surface geometry, and 3D mapping from a single architecture.
OpenMOSS releases MOSS-Audio — an open audio model that outperforms larger alternatives
OpenMOSS introduced MOSS-Audio — an open model that understands speech, music, and ambient sounds in a single architecture and outperforms significantly larger open-source models on common audio benchmarks.
OpenAI Embeddings and RL: How to Build an Agent with Long-Term Memory for Accurate Answers
The tutorial shows how to train an RL agent to select relevant records from long-term memory so that an LLM answers questions about saved facts more accurately.
How to Measure Real Intelligence: Key Benchmarks for AI Agents
Classical tests no longer reflect the real capabilities of neural networks. We explore which benchmarks truly show whether an AI agent is ready for real-world business applications.
Elastic Memory for AI: How kvcached Solves the GPU Shortage
Dynamic KV-cache distribution promises to radically reduce the cost of hosting language models by enabling efficient memory sharing across a single GPU.
xAI's Voice Model Surpasses GPT Realtime in Business Tasks
Elon Musk's xAI has unveiled grok-voice-think-fast-1.0. The new model outperformed solutions from OpenAI and Google in managing complex business workflows.
GitNexus: How a New Tool Taught AI Agents Structural Code Understanding
AI assistants excel at writing local code but often fail to grasp the global architecture of projects. The open-source GitNexus project solves this problem using knowledge graphs and the MCP protocol.
DeepSeek-V4: How New Compression Algorithms Made One-Million-Token Context a Reality
DeepSeek introduced the fourth generation of its models. Through radical attention compression, processing massive data volumes is becoming accessible not just to tech giants.
Decoupled DiLoCo Architecture from DeepMind Solves AI Scaling Problem
Training advanced AI models has always been held hostage by hardware failures: one chip breaking would stop the entire cluster. A new architecture from Google DeepMind changes the rules of the game.
OpenMythos: Open-source PyTorch reconstruction of Claude Mythos architecture with 770M parameters
Developer Kye Gomez reconstructed the presumed Claude Mythos architecture from Anthropic from scratch — without leaks, based only on public research.
OpenAI Scales GPT-5.4-Cyber for Verified Cybersecurity Professionals
OpenAI opens broad access to GPT-5.4-Cyber — a fine-tuned version of GPT-5.4 without standard restrictions for thousands of verified defenders.
Moonshot AI releases Kimi K2.6: an agentic model with a swarm of 300 sub-agents
Chinese lab Moonshot AI has open-sourced Kimi K2.6, a multimodal agentic model that autonomously coordinates up to 300 sub-agents and 4,000 steps.
Microsoft Phi-4-Mini: implementing quantization, RAG, and LoRA in a single Jupyter notebook
The tutorial shows the full Phi-4-mini-instruct pipeline: 4-bit quantization, streaming generation, reasoning, tool calling, RAG, and LoRA fine-tuning in one notebook.
Qwen 3.6-35B-A3B in practice: multimodality, MoE, and RAG in a single pipeline
A detailed breakdown of implementing Qwen 3.6-35B-A3B, from model loading to RAG, tool calling, and session persistence in practical workflows.
AI Agents Now Write GPU Code: RightNow AI Releases the AutoKernel Framework
Writing fast GPU code is notoriously hard. RightNow AI wants to automate it entirely with AutoKernel, a new open-source framework built on AI agents.
Evolving AI Agents: A Step-by-Step Guide to the A-Evolve Framework
Learn how to build and configure an evolving OpenAI-based AI agent with the A-Evolve framework. A hands-on guide to memory, skills, and the agent workspace.
Tutorial: Building Cybersecurity AI Agents with the CAI Framework
A step-by-step guide to building advanced cybersecurity AI agents: set up multi-agent systems, custom tools and guardrails with the CAI framework.
NVIDIA introduced Nemotron 3 Super — an open model with 120 billion parameters
The new hybrid model combines Mamba and Attention architectures with a Mixture-of-Experts approach, delivering a fivefold performance gain for multi-agent systems. The gap between proprietary and open models is rapidly n
Google introduced TensorFlow 2.21 and LiteRT for mobile AI
Google introduced TensorFlow 2.21 and LiteRT, a new standard for running neural networks on smartphones with NPU support and GPU acceleration for mobile platforms.
Google launches Android Bench to evaluate AI in mobile development
Google has made Android Bench available, the first specialized tool for assessing large language models' skills in Android app development.
OpenAI unveils Codex Security for automated detection and remediation of code vulnerabilities
OpenAI is launching Codex Security, an intelligent agent for code security analysis that not only finds issues, but also suggests ready-made patches to fix them.
Liquid AI released a system for running AI agents fully on-device
Liquid AI unveiled the LFM2-24B-A2B model and the LocalCowork app — a pairing for executing complex AI agent workflows without a single call to the cloud. Privacy is becoming an architectural principle.
Yuan 3.0 Ultra: one trillion parameters with record efficiency
China's YuanLab AI has unveiled an open multimodal model with one trillion parameters, while activating only 68.8 billion of them. A one-third reduction in parameters alongside a 49% increase in training efficiency — how
Alibaba releases OpenSandbox, a unified environment for safe AI agent operations
Alibaba has open-sourced OpenSandbox, a tool that gives AI agents isolated sandboxes for code execution, web browsing, and model training. It is a bid to set a standard for the entire agentic AI infrastructure.
Alibaba released Qwen 3.5 Small — compact models that run directly on devices
Alibaba’s new model series, ranging from 0.8 to 9 billion parameters, challenges the race for size. The philosophy of «more intelligence, less compute» could reshape mobile AI.
SHAP-IQ: a new standard for explainable AI enters practical use
The SHAP-IQ library makes it possible not only to assess the importance of a model’s features, but also to understand how they interact with one another. We explain why this changes the approach to interpreting machine l
Google introduces STATIC: 948x faster generative search
Google AI developed a sparse-matrix-based framework that speeds up decoding in generative recommendation systems by nearly 1,000x. This could change the approach to industrial-scale recommendations.
Alibaba unveils CoPaw — a workstation for scaling AI agents
Alibaba has released the open-source framework CoPaw, which turns a developer's local machine into a full-fledged environment for running autonomous AI agents with memory and multichannel workflows.
Unified Latents: Google DeepMind finds a way to improve AI generation
Google DeepMind's new Unified Latents framework solves the dilemma between data compression quality and neural network accuracy.
Sakana AI has learned to instantly adapt language models without fine-tuning
Tokyo-based lab Sakana AI introduced two methods — Doc-to-LoRA and Text-to-LoRA — that make it possible to adapt large language models to new data in fractions of a second, without classical fine-tuning.