MarkTechPost
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
The Complete Guide to ModelScope: From Model Discovery to Export
A detailed hands-on guide to working with the ModelScope platform in Google Colab. Learn how to search for, test, fine-tune and export AI models in a streamlined workflow.
OSGym: A New Framework for Training AI Agents at Scale on the Cheap
OSGym is an infrastructure framework that runs thousands of virtual environments for training AI to operate a computer, at roughly the price of a cup of coffee per day.
Meta Unveils Muse Spark: A Multimodal AI Model With Visual Reasoning
Meta Superintelligence Labs has introduced Muse Spark, the first AI model in a new family. It supports visual chain of thought and works with parallel agents. *Meta is designated an extremist organization and banned in
Sigmoid vs. ReLU: The Hidden Cost of Losing Geometric Context
Neural networks behave like complex geometric systems. Why preserving spatial information matters for inference, and how the choice between Sigmoid and ReLU changes the rules.
Building AI Document-Processing Pipelines with LangExtract and OpenAI
A new guide walks through building pipelines that extract structured data from text using Google's LangExtract and OpenAI language models.
Google Unveils PaperOrchestra: An AI System That Writes Scientific Papers Automatically
Researchers have built PaperOrchestra, a multi-agent system that automates the hard work of turning rough notes and experimental data into a finished scientific paper.
How to Build a Stable Pipeline on the Fable 5 Traces Dataset in Google Colab
A step-by-step tutorial on working with the Fable 5 Traces dataset: manual JSONL parsing, tool call normalization, secret auditing, and training a basic classifier in Colab without heavy dependencies.
6 No-Code Tools for AI Developers: From RAG to Multi-Agent Systems
Six no-code platforms, including Flowise, Dify and Stack AI, let developers build RAG systems, multi-agent workflows and fine-tune LLMs without writing code.
Thinking Machines Lab Releases Inkling-Small: Open 276B MoE Model Outperforms Its Teacher
Thinking Machines Lab has open-sourced the weights of Inkling-Small, a 276B-parameter multimodal MoE model. At a quarter of its teacher's size, it beats it on SWE-bench, HLE and ARC-AGI-2.
NVIDIA Releases Molt: A PyTorch Framework for Agentic RL in 8,600 Lines of Code
NVIDIA NeMo has released Molt, a compact PyTorch framework for agentic RL: 8,600 lines of code versus 62,000 for competitors, with performance comparable to the Megatron stack.
NVIDIA Transformer Engine: Fused Kernels, FP8 and BF16 for Training Transformers
NVIDIA Transformer Engine speeds up transformer training with fused GPU kernels and FP8/BF16 formats. A tutorial shows how to plug TE into a GPT-style model and measure the memory savings.
Tencent Open-Sources AngelSpec, a Framework for Speeding Up LLM Inference With Speculative Decoding
Tencent has open-sourced AngelSpec, a torch-native framework for training speculative decoding draft models that speeds up LLM generation without lowering answer quality.
Moonshot AI Open-Sources MoonEP, the Library Used to Train the MoE Model Kimi K3
Moonshot AI released MoonEP, an expert-parallelism library, under the MIT license. It gives every GPU rank exactly S×K tokens and sped up Kimi K3 scaling by 2.5x.
Prompt vs Loop vs Graph Engineering: Three Levels of Control in AI Development
On July 29, 2026, MarkTechPost laid out prompt, loop and graph engineering as three nested levels of control rather than competing approaches.
Token Saver: open-source MCP extension cuts Claude's token use on PDFs by 90-99%
Marktechpost has released Token Saver, a free MCP extension for Claude Desktop that processes large PDFs locally and cuts token use by 92-99%.
Fireworks Nexus: Routing Code Tasks to Open-Weight Models and Cutting AI Costs 3–5x
Fireworks Nexus sends routine code to cheaper open-weight models and cuts AI agent costs 3–5x, without changing the tools developers already use.
Microsoft Releases MAI-Cyber-1-Flash: 5B Active Parameters and 95.95% on CyberGym
Microsoft AI has released MAI-Cyber-1-Flash, a 5B-active-parameter cyberdefense model that, paired with the MDASH harness, lifted its CyberGym score to 95.95%.
Liquid AI Releases LFM2.5-Encoder-230M and 350M Encoders With an 8K-Token Context
Liquid AI has released open LFM2.5-Encoder-230M and 350M encoders with an 8192-token context. They stay fast on long inputs even on a CPU and compete with models many times their size.
Perplexity Releases pplx: A Search API CLI for the Terminal at $5 per 1,000 Requests
Perplexity has launched pplx, an official CLI that returns Search API results to the terminal as JSON: two commands, $5 per 1000 requests, built for coding agents.
Black Forest Labs releases FLUX 3: up to 20 seconds of video with sound from a single model
Black Forest Labs unveils FLUX 3, a multimodal model that generates images, up to 20 seconds of video with sound, and robot actions from one set of weights.
Claude Sonnet 5 vs Opus 4.8: agentic coding comparison and API pricing from Anthropic
MarkTechPost compared Claude Sonnet 5, Sonnet 4.6, and Opus 4.8 from Anthropic: the new model approaches the flagship in agentic coding at lower Sonnet token prices.
Linq Launched iMessage Apps — Payments, Tickets and Games for AI Agents Directly in Chats
Linq introduced iMessage Apps — interactive imessage_app cards for payments, ticket purchases, flight bookings and games within messaging.
Google Research Releases TabFM — Foundation Model for Tabular Data with Zero-Shot
Google Research presented TabFM — a foundation model for tables that performs zero-shot classification and regression without training for each dataset.
Baidu Develops CUP — Set of Python Tools for Reliable Backend Pipelines
MarkTechPost analyzed CUP — a set of Python utilities from Baidu for logging, caching, thread pools and Linux resource monitoring in one package.
Sakana AI Released Sakana Marlin — an Agent for 100-Page Reports with Slides
Sakana AI Introduced Its First Commercial Product — the Corporate Agent Marlin, Which Works Autonomously for Up to Eight Hours and Generates Reports Up to 100 Pages with Slides
Meta AI Presented Brain2Qwerty v2 — Non-invasive MEG Text Decoder with 61% Accuracy
Meta AI released Brain2Qwerty v2 — a non-invasive MEG-based pipeline that decodes typed sentences from brain signals with 61% word-level accuracy and published the training code.
Anthropic resumed Claude Fable 5 operations after US export restrictions were lifted
Anthropic on July 1, 2026 restored access to Claude Fable 5 and introduced a classifier that blocks jailbreak techniques.
OpenAI Released GPT-Realtime-2.1 and Mini Version for Voice Agents in API
OpenAI added two models for voice agents to its API—GPT-Realtime-2.1 and GPT-Realtime-2.1-mini—reducing p95 latency by at least 25% through caching.
Tencent released Hy3—an open MoE model with 295 billion parameters and 256K context
Tencent presented Hy3—a model with 295 billion parameters, 21 billion active per token, Apache 2.0 license and a 78.0 score on SWE-Bench Verified.
MarkTechPost showed how to build an AI co-researcher QSAR for finding EGFR inhibitors
A MarkTechPost tutorial teaches how to build an autonomous AI co-researcher using Random Forest QSAR — from ChEMBL and RDKit data to ready-made candidate molecules against EGFR C797S.