MarkTechPost
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
WebBrain: A Browser AI Agent That Runs Without the Cloud
WebBrain is a free, MIT-licensed agent for Chrome and Firefox that reads pages and automates tasks, running locally through llama.cpp or Ollama.
xAI launches /goal in Grok Build: an autonomous agent that plans and verifies multistep tasks
Grok Build's new /goal mode takes a single goal, builds its own plan, works through a checklist, and verifies the result at every stage until the task is done.
MoonMath AI releases open-source HIP kernel for AMD MI300X that outpaces official AITER v3
Startup MoonMath AI has open-sourced an attention kernel for the AMD MI300X that beats AMD's official library across all tensor shapes and rounding modes.
Cisco open-sources FAPO: automatic LLM pipeline optimization with diagnosis at every step
Cisco Foundation AI has open-sourced FAPO, a Claude Code-based tool that finds the weak step in a prompt chain on its own and tunes the pipeline to a target accuracy.
TimeCopilot: Building a Forecasting Pipeline with Foundation Models and Anomaly Detection
TimeCopilot covers the full forecasting cycle, from statistical to foundation models, with automatic anomaly detection and an LLM agent that explains its predictions.
Docling Parse: A PDF Parsing Pipeline Accurate to the Character and Coordinate
A tutorial shows how to use Docling Parse for structural PDF analysis: coordinates for every word, visual overlays, and JSON and CSV export for RAG systems.
MiniMax releases MSA: sparse attention speeds up million-token context processing by 28x
MiniMax unveiled MSA, a two-branch sparse attention mechanism trained on a 109-billion-parameter MoE model that runs 28 times faster than standard GQA at a 1 million token context.
Cohere's North Mini Code: An Open 30B MoE Model for Agentic Coding
Cohere has released its first open-weight developer model, North Mini Code: a 30B-parameter MoE with a 256K context window that runs on a single H100 GPU.
Microsoft SkillOpt: Automatic Prompt Optimization Instead of Manual Trial and Error
Microsoft SkillOpt optimizes AI prompts through an automated six-step loop of reflection, selection and validation gating, and shows measurable accuracy gains within the first few iterations.
Google DeepMind Releases DiffusionGemma, a 26B MoE Model That Generates Text Up to 4x Faster
Google DeepMind has released DiffusionGemma, an experimental open 26B-parameter model that uses text diffusion instead of autoregression, speeding up generation by up to 4x on GPUs.
PyGraphistry: Graph Analytics for Threat Detection in Enterprise Data
Researchers have published a PyGraphistry Colab workflow that turns enterprise access data into an interactive graph with ML anomaly scoring and three visualization modes for incident investigation.
Alibaba launches Qwen3.7-Plus: a multimodal agent that analyzes video and writes code on its own
Alibaba has released Qwen3.7-Plus, a multimodal agent on its Bailian platform. It sees images and video, calls tools, and iterates on code without an operator.
JetBrains releases Mellum2, an open 12B MoE model for fast tasks in AI pipelines
JetBrains has open-sourced Mellum2 under Apache 2.0: a compact 12-billion-parameter MoE model trained on 10.6 trillion tokens, built to slot into multi-model AI pipelines.
How to Speed Up Transformer Training with NVIDIA Apex: FusedAdam, FusedLayerNorm and torch.amp
A step-by-step guide to building NVIDIA Apex from source and using FusedAdam, FusedLayerNorm and mixed precision, for up to 2.5× faster transformer training.
Anthropic releases Claude Fable 5 and Mythos 5: one model, two access tiers
Fable 5 is public with safety classifiers; Mythos 5, with cyber restrictions lifted, is available only through the closed Project Glasswing program.
How to Build an AI Agent in Google Colab Without Frameworks: A Tutorial With Memory and MCP
A step-by-step tutorial for building a lightweight AI agent in Google Colab with tool registration, session memory, skills, and an MCP server, all without external frameworks.
OpenAI unveils GPT-5.6 with three tiers, Sol, Terra and Luna, and new reasoning modes
OpenAI has announced the GPT-5.6 family of three models, Sol, Terra and Luna, plus new max and ultra reasoning modes. Access is limited to selected developers for now.
Perplexity launches Computer for Counsel, an AI agent for legal teams
Perplexity has launched an AI agent built for lawyers: it routes queries across 20+ models, integrates with Microsoft 365 and Midpage, and backs every answer with verifiable source links.
NVIDIA Open-SWE-Traces: Building a Dataset to Fine-Tune Coding Agents
A tutorial shows how to work with NVIDIA's Open-SWE-Traces dataset: parse agent trajectories, analyze code patches and assemble an SFT set for training AI agents.
DeepSeek releases DSpark, a framework that speeds up DeepSeek-V4 generation by 57–85%
DeepSeek has open-sourced DSpark, a speculative decoding framework for DeepSeek-V4 that speeds up generation by 57–85% with no loss in quality.
Meta open-sources Astryx, a React design system built for AI agents and MCP
Meta has released Astryx, an open-source design system on React and StyleX with an MCP server and CLI, where engineers and AI agents work through one API with the same components. *Meta is designated an extremist organiz
Liquid AI releases LFM2.5-230M: 213 tokens/s on Galaxy S25 and llama.cpp support
Liquid AI has released LFM2.5-230M, an open 230M-parameter model that runs offline on a smartphone and a Raspberry Pi, beating rivals four times its size at instruction following.
Datalab releases lift, an open-weights 9B model that extracts JSON from PDFs
Datalab has open-sourced lift, a 9-billion-parameter vision model that turns PDFs and images into structured JSON following a given schema, scoring 90.2% field accuracy on a 225-document benchmark.
Z.ai releases GLM-5.2: a truly usable 1 million tokens and two levels of deep thinking
Z.ai has launched GLM-5.2 with a genuinely working 1M-token context, High and Max modes, and support for Claude Code, Cline and OpenClaw via an Anthropic-compatible API.
Training Datasets for Neural Networks: How to Use FineWeb (15 Trillion Tokens)
A practical guide to Hugging Face's FineWeb: how to work with a multi-terabyte web corpus for LLM training through streaming, filtering, deduplication and tokenization, without downloading the full data.
Zyphra releases Zamba2-VL: vision-language models with 10x faster response
Zyphra has released Zamba2-VL, open VLMs at 1.2B, 2.7B and 7B parameters. Their hybrid Mamba2 + Transformer architecture cuts time to first token by roughly 10x.
Moonshot AI launches Kimi Work, a desktop agent that runs a swarm of 300 subagents
Moonshot AI's Kimi Work runs locally on macOS and Windows, drives the browser through WebBridge, and coordinates a swarm of up to 300 parallel subagents.
Claude Code: An AI IDE for Developers Instead of ChatGPT
MarkTechPost has published a guide to Claude Code 2026 covering 25 features, from CLAUDE.md and skills to MCP, hooks and Auto Mode, with code examples and a demo.
A Neural Network for SQL: Google Improves Gemini 3.1 Pro
Google Research unveils Gemini-SQL2, built on Gemini 3.1 Pro. It scored 80.04% accuracy on the BIRD benchmark for converting text into SQL queries.
QwenPaw: What It Is and How to Build an Agent Workspace
A step-by-step tutorial on deploying QwenPaw, a framework for building AI agents with multiple model providers, custom skills and a streaming API, in Google Colab.