MarkTechPost
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Liquid AI releases LFM2.5-350M: a 350M-parameter model trained with large-scale RL
Liquid AI argues that intelligence density matters more than parameter count: a small model can be trained to perform better through reinforcement learning.
Parallax: Local Linear Attention with Covariance Correction
MarkTechPost reports on Parallax, a new attention variant for small LLMs that claims better quality and faster decoding.
MiniMax M3: a new model with 1M-token context, multimodality and computer use
MiniMax has unveiled MiniMax Sparse Attention, an architecture that supports a million-token context and handles video, images and computer control.
Autonomous AI agents run about 50 times longer than search assistants, Harvard and Perplexity study finds
A new Harvard and Perplexity study finds AI agents work on their own for 26 minutes per session, versus 33 seconds for search assistants.
Memory OS: An Open-Source, Six-Layer Memory Stack for Hermes Agent
The new Memory OS project gives Hermes Agent local, persistent memory through six architectural layers, gated retrieval and a built-in wiki.
Google launches Gemini 3.5 Live Translate: a speech-to-speech model for 70+ languages
Gemini 3.5 Live Translate translates speech in real time, generating audio a few seconds behind the speaker. It is available in the Gemini Live API, Google Meet and the Google Translate app.
Nous Research unveils a profile builder for Hermes Agent
The new Hermes Agent interface replaces multistep CLI setup: users now build an agent profile in a single flow on the dashboard.
SpaceXAI releases Grok 4.5, an Opus-class model built for coding
SpaceXAI has unveiled Grok 4.5, an Opus-class model trained with Cursor for coding and agentic tasks, priced at $2/$6 per million tokens.
Ant Group unveils LingBot-VLA 2.0, an open model for controlling robots of any design
On July 8, 2026, Robbyant (Ant Group) released LingBot-VLA 2.0, an open 6-billion-parameter model that controls robots of different designs under the Apache 2.0 license.
NVIDIA Nemotron-Labs-3-Puzzle-75B: 37% Smaller Model Doubles Server Throughput
NVIDIA has released a compressed version of its Nemotron-3-Super LLM. Using the Puzzle method, which alternates compression and distillation, the model is 37% smaller and doubles server throughput.
Google Research releases SensorFM, a model for 35 health prediction tasks
Google Research trained the SensorFM foundation 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 has unveiled LingBot-World-Infinity, a 14B video model that works as an interactive simulator and generates 720p video at 60 frames per second.
AI for documents: Datalab Lift vs. the competition
Datalab compared its 9B Lift tool with NuExtract3, LlamaExtract, Marker, and Docling, and explained when a schema-first approach wins at extracting data from PDFs.
Liquid AI releases Antidoom: FTPO method cuts hangs in reasoning models
Liquid AI has open-sourced Antidoom, a tool that finds the token that triggers an endless loop and retrains only that one, cutting the doom-loop rate from 22.9% to 1%.
Google Gemini: Import GitHub Repositories into Build Mode
Google AI Studio has updated Build mode: any GitHub repository can now be imported directly and turned into an editable, deployable app right away.
Audex-30B: NVIDIA's Neural Network for Audio and Speech
NVIDIA has combined speech recognition, translation, synthesis and audio generation in a single 30-billion-parameter MoE model that retains the text abilities of Nemotron-Cascade-2.
LingBot-Vision: Ant Group's Free Vision Model
Ant Group's Robbyant has released LingBot-Vision, a 1B-parameter ViT trained to recognize object boundaries without labels, matching or beating larger models on spatial perception tasks.
AI for PDF Documents: A Guide to Tools in 2026
A survey of open-source tools that convert corporate PDFs and scans into structured JSON, without which LLMs and AI agents cannot work with real enterprise data.
Meituan releases LongCat-2.0: an open 1.6-trillion-parameter MoE model with a 1-million-token context
Chinese company Meituan has opened up LongCat-2.0, a giant MoE model with a native 1-million-token window, trained on domestic ASICs.
Synthetic Sciences releases OpenScience, an open AI workbench for ML, biology, physics and chemistry
OpenScience from Synthetic Sciences is an open-source workbench under Apache 2.0 with 250+ skills, compatible with any AI model for ML, biology, physics and chemistry.
Sakana AI launches Sakana Translate for Japanese, English and Chinese
Sakana AI has added Sakana Translate, built on its Namazu models, to Sakana Chat, with three modes: direct translation, proofreading and a dialogue with the document.
How to Train Gemma-3 on Math with 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 exporting the final model.
Alibaba's Page Agent Controls Web Interfaces Through the DOM, With No Screenshots or Multimodal Models
Alibaba has introduced Page Agent, a client-side JavaScript agent that takes text commands and drives web pages through the DOM, with no screenshots and no server-side code.
Anthropic launches Claude Science Beta: a 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 unveils ASPIRE, a self-improving robotics framework that hits 31% zero-shot on long-horizon tasks
NVIDIA released ASPIRE, a robotics framework that writes and refines control programs, stores verified solutions in a skill library, and applies them to new tasks without fine-tuning.
NVIDIA HORIZON: Autonomous Agent Completes 100% of RTL Benchmarks in Chip Design
NVIDIA has unveiled HORIZON, an agent that writes and verifies RTL chip-design code on its own, reaching 100% completion on standard benchmarks.
Former Qwen tech lead: why hybrid thinking didn't work and what comes next
Junyang Lin, former Qwen lead at Alibaba, explains where Qwen3's dynamic thinking fell short and why agentic AI demands a fundamentally different approach to training.
AI Document Search: LlamaIndex Releases legal-kb for Free
LlamaIndex has published legal-kb, a reference app that gives agents file-system-style access to a knowledge base via retrieve, find, read and grep on Index v2.
NVIDIA releases Nemotron-Labs-TwoTower, an open-weight text model built on diffusion
NVIDIA has opened the weights of Nemotron-Labs-TwoTower, a diffusion language model that pairs a frozen AR backbone with a diffusion head to speed up text generation.
Interfaze releases open diffusion-based ASR model covering six languages
Interfaze has published diffusion-gemma-asr-small, the first open ASR adapter built on Google's DiffusionGemma: 42M parameters, six languages, inference cost set by the number of denoising steps.