MarkTechPost
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Liquid AI released LFM2.5-350M: efficient 350M parameter model with scaled RL
Liquid AI proved that AI density is more important than parameter count: small model can be trained better through reinforced learning.
Parallax: improved local linear attention with covariance correction
Researchers from MarkTechPost presented Parallax — a new variation of the attention mechanism for small LLMs with improved quality and speed.
MiniMax M3: New Model with 1M Tokens, Multimodality, and Computer Use
The company introduced the MiniMax Sparse Attention architecture supporting one million tokens in context and work with video, images, and computer control.
Autonomous AI agents run 50x longer than search: Harvard and Perplexity research
New research from Harvard and Perplexity shows AI agents perform 26 minutes of independent work per session, compared to 33 seconds for search assistants.
Memory OS: Open 6-Layer Memory Stack for Hermes Agent
New Memory OS project adds persistent local memory to Hermes Agent through six architecture layers, managed memory search, and embedded wiki.
Google launches Gemini 3.5 Live Translate: speech-to-speech model for 70+ languages
Gemini 3.5 Live Translate translates speech in real time, synthesizing audio with a lag of several seconds. The model is available in Gemini Live API, Google Meet, and Google Translate app.
Nous Research presented Hermes Agent profile builder
New Hermes Agent interface replaces multi-step CLI configuration — users now assemble agent profiles in a single dashboard flow.
SpaceXAI Released Grok 4.5 — Opus-Class Code Model
SpaceXAI presented Grok 4.5, Opus-class model trained with Cursor for coding and agentic tasks, priced at $2/$6 per million tokens.
Ant Group Presented LingBot-VLA 2.0 — Open Model for Controlling Robots of Any Design
Its 6-billion parameter model trained on 60 thousand hours of robotics video and can control a dozen types of manipulators through a unified action space.
NVIDIA Nemotron-Labs-3-Puzzle-75B: 37% model compression doubled server speed
NVIDIA released a compressed version of the Nemotron-3-Super LLM — Nemotron-Labs-3-Puzzle-75B. Thanks to the Puzzle method (alternating compression and distillation), the model became 37% smaller and doubled server throu
Google Research released SensorFM — a model for 35 health forecasting tasks
Google Research trained the foundational SensorFM model on 1 trillion minutes of data from 5 million Fitbit and Pixel Watch users to predict health outcomes.
Robbyant releases LingBot-World-Infinity, a causal model for interactive worlds
Robbyant unveiled LingBot-World-Infinity, a 14B video model that works as an interactive simulator, generating 720p video at 60 frames per second.
Datalab Lift vs Competitors: How a 9B Document Extractor Works with JSON Schema
Datalab compared its 9B Lift tool with NuExtract3, LlamaExtract, Marker, and Docling — and analyzed when the schema-first approach wins in PDF data extraction.
Liquid AI Released Antidoom: FTPO Method Reduces Hangs in Reasoning Models
Liquid AI open-sourced Antidoom — a tool that identifies the token trigger of infinite loops and retrains only it, reducing doom-loop frequency from 22.9% to 1%.
Google AI Studio added GitHub repository import to Build development mode
Google AI Studio updated Build mode — any GitHub repository can now be imported directly and immediately turned into an editable, deployable app.
NVIDIA released Audex-30B — a unified audio-text MoE model based on Nemotron-Cascade-2
NVIDIA combined speech recognition, translation, speech synthesis, and audio generation in a single 30-billion-parameter MoE model while preserving Nemotron-Cascade-2’s text capabilities.
Ant Group open-sources LingBot-Vision: a 1B model for robot spatial perception
Ant Group's Robbyant has released LingBot-Vision, a 1B ViT model trained to detect object boundaries without labels and outperform larger counterparts on spatial perception tasks.
AI for PDF Document Processing: 2026 Tools Guide
Overview of open-source tools for converting corporate PDFs and scans into structured JSON — essential for LLMs and AI agents to work with real corporate data.
Meituan Releases LongCat-2.0: Open MoE Model with 1.6 Trillion Parameters and 1 Million Token Context
Chinese company Meituan unveiled LongCat-2.0 — a massive MoE model with native one-million-token context window trained on domestic ASICs.
Synthetic Sciences Releases OpenScience — Open AI Workbench for ML, Biology, Physics and Chemistry
OpenScience by Synthetic Sciences — open-source workbench under Apache 2.0 with 250+ skills, compatible with any AI model for ML, biology, physics and chemistry tasks.
Sakana AI Launches Sakana Translate Translator for Japanese, English, and Chinese
Sakana AI added Sakana Translate translator to Sakana Chat, built on Namazu models — with three modes: direct translation, proofreading, and document dialogue.
Fine-tuning Gemma-3 on Mathematics Using GRPO and LoRA Adapters
A step-by-step GRPO workflow for fine-tuning Gemma-3 on the GSM8K dataset—from environment setup and LoRA adapters to reward functions and final model export.
Page Agent from Alibaba Manages Web Interfaces via DOM Without Screenshots and Multimodal Models
Alibaba introduced Page Agent — a client-side JavaScript agent that understands text commands and manages web pages via DOM without screenshots and server-side code.
Anthropic Launches Claude Science Beta: Multi-Agent Environment for Genomics and Proteomics
On June 30, 2026, Anthropic opened beta access to Claude Science — a multi-agent platform for reproducible pipelines in genomics, proteomics, and cheminformatics.
NVIDIA Introduces ASPIRE — Self-Learning Robotics Framework Achieving 31% Zero-Shot Success on Complex Tasks
NVIDIA released ASPIRE — a robotics framework that writes and improves control programs, accumulates verified solutions in a skill library, and transfers them to new tasks without fine-tuning.
NVIDIA HORIZON: Autonomous Agent Closes 100% of RTL Benchmarks in Chip Design
NVIDIA introduced HORIZON — an agent that independently writes and verifies RTL code for chip design, achieving 100% completion on standard benchmarks.
Former Qwen Technical Lead: Why Hybrid Reasoning Failed and What's Next
Zunyuan Lin, former Qwen development lead at Alibaba, explained where Qwen3's dynamic reasoning failed—and why agentic AI requires a fundamentally different training approach.
LlamaIndex released legal-kb: agentic document search through four tools on Index v2
LlamaIndex published the legal-kb reference application, giving agents file-like access to a knowledge base via retrieve, find, read, and grep on the Index v2 engine.
NVIDIA Released Nemotron-Labs-TwoTower — Diffusion Language Model with Open Weights
NVIDIA has opened the weights of the diffusion language model Nemotron-Labs-TwoTower, which combines a frozen AR backbone with a diffusion head to accelerate text generation.
Interfaze released an open-source diffusion ASR model for speech recognition in six languages
Interfaze published diffusion-gemma-asr-small — the first open-source ASR adapter built on Google's DiffusionGemma: 42 million parameters, six languages, and inference cost tied to the number of denoising steps.