MarkTechPost
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Meta unveils Sapiens2, a single vision model for pose, segmentation and 3D
Meta Reality Labs has released Sapiens2, a family of human-focused vision models that handles pose, segmentation, surface geometry and 3D maps with one architecture.
OpenMOSS releases MOSS-Audio, an open audio model that beats larger rivals
OpenMOSS has unveiled MOSS-Audio, an open model that handles speech, music and ambient sound in a single architecture and outperforms open-source models noticeably larger than itself on general audio benchmarks.
OpenAI Embeddings and RL: How to Build an Agent With Long-Term Memory for Accurate Answers
A tutorial shows how to train an RL agent to pick relevant entries from long-term memory so an LLM can answer questions about stored facts more accurately.
How to Measure Real Intelligence: The Key Benchmarks for AI Agents
Classic tests no longer reflect what neural networks can actually do. We look at which benchmarks show whether an AI agent is ready for real-world business work.
Elastic Memory for AI: How kvcached Tackles the GPU Shortage
Dynamic KV cache allocation promises to cut the cost of hosting language models sharply by letting a single GPU's memory be shared efficiently.
xAI's voice model beats GPT Realtime on business tasks
Elon Musk's xAI has unveiled grok-voice-think-fast-1.0. The new model outperformed OpenAI and Google offerings at managing complex business workflows.
GitNexus: A New Tool Teaches AI Agents to Understand Code Structure
AI assistants write good local code but often miss a project's global architecture. The open-source project GitNexus tackles this with knowledge graphs and the MCP protocol.
DeepSeek-V4: How New Compression Algorithms Made a One-Million-Token Context Real
DeepSeek has unveiled the fourth generation of its models. Radical compression of attention puts huge data volumes within reach of more than just the tech giants.
DeepMind's Decoupled DiLoCo Tackles the AI Scaling Problem
Training frontier AI models has always been hostage to hardware failures: one broken chip could halt an entire cluster. A new architecture from Google DeepMind changes that.
OpenMythos: an open-source PyTorch reconstruction of the Claude Mythos architecture with 770M parameters
Developer Kye Gomez rebuilt the alleged Claude Mythos architecture from scratch, without leaks, using only public research.
OpenAI scales GPT-5.4-Cyber for verified cybersecurity professionals
OpenAI is opening broad access to GPT-5.4-Cyber, a fine-tuned GPT-5.4 without standard restrictions, to 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 coordinates up to 300 sub-agents and 4,000 steps autonomously.
Microsoft's Phi-4-Mini: Quantization, RAG and LoRA in a Single Jupyter Notebook
A tutorial walks through a full pipeline on Phi-4-mini-instruct: 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
On April 21, 2026, MarkTechPost published a tutorial on the multimodal MoE model Qwen 3.6-35B-A3B, covering GPU loading, MoE routing, RAG and session persistence.
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 unveils Nemotron 3 Super, an open 120B-parameter model
The new hybrid model pairs Mamba and Attention with Mixture-of-Experts for a fivefold throughput gain in multi-agent systems. The gap between proprietary and open models is closing fast.
Google releases TensorFlow 2.21 and LiteRT for mobile AI
Google has released 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 opened access to Android Bench, the first specialized tool for testing how well large language models handle Android app development.
OpenAI unveils Codex Security to find and fix code vulnerabilities automatically
OpenAI is launching Codex Security, an AI agent for code security analysis that not only finds flaws but also proposes ready-made patches to fix them.
Liquid AI releases a system for running AI agents entirely on-device
Liquid AI has unveiled LFM2-24B-A2B and LocalCowork, a pairing built to run complex AI-agent workflows without a single call to the cloud. Privacy becomes an architectural principle.
Yuan 3.0 Ultra: One Trillion Parameters, Record Efficiency
China's YuanLab AI has released an open multimodal model with a trillion parameters that activates only 68.8 billion of them. Parameters cut by a third, training efficiency up 49%: how does it work?
Alibaba releases OpenSandbox, a unified environment for running AI agents safely
Alibaba has open-sourced OpenSandbox, a tool that gives AI agents isolated sandboxes for running code, browsing the web and training models. It is a bid to set the standard for agentic AI infrastructure.
Alibaba releases Qwen 3.5 Small, compact models built to run directly on devices
Alibaba's new model series, from 0.8 to 9 billion parameters, pushes back against the race for size. Its "more intelligence, less compute" approach could change the rules for mobile AI.
SHAP-IQ: A New Standard for Explainable AI Moves Into Practical Use
The SHAP-IQ library goes beyond ranking feature importance and shows how a model's features interact with each other. Here is why that changes how machine learning gets interpreted.
Google unveils STATIC, a 948x speedup for generative retrieval
Google AI has built a sparse-matrix framework that speeds up decoding in generative recommender systems by almost a thousand times. It could change how recommendations are run in production.
Alibaba open-sources CoPaw, a workstation for scaling AI agents
Alibaba has released CoPaw, an open-source framework that turns a developer's local machine into a full 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 tackles the trade-off between data compression quality and neural network accuracy.
Sakana AI Can Now Adapt Language Models Instantly, Without Fine-Tuning
Tokyo-based lab Sakana AI has unveiled Doc-to-LoRA and Text-to-LoRA, two methods that adapt LLMs to new data in a fraction of a second, without classic fine-tuning.