90 days of the enterprise, pro and team plans free, no card needed. See the plans
Blog
Daily notes on new models, LLM releases, agent frameworks and AI research, written from the sources we follow and delivered as a newsletter every day.
Get the morning and evening issues
Every new post of the day, in one email. Confirmation required.
AI13 min read
NVIDIA launched Nemotron 3.5 Bolt E4B, a small language model running on 4GB of VRAM. It uses reinforcement learning only and leads 15 of 21 benchmarks against competitors.
From MarkTechPost
AI1 min read
OpenAI released 719 math manuscripts from an unreleased frontier model. Mistral and Reflection AI launched new open-weight models to compete with Chinese systems.
From Last Week in AI
Photo: Yan Krukau
AI10 min read
Reflection released Beam, an open-weight model that matches GLM 5.2 and Qwen 3.8 on benchmarks while using three to four times less compute.
From The Decoder
AI9 min read
Reka released a research preview of Rho-1, a 19B model that generates video and outputs robot actions. It replaces pipelines by handling all tasks within one context window.
From MarkTechPost
Research1 min read
Apple researchers found diffusion models are often overconfident in discrete tasks. They also warn that on-policy distillation can hurt reasoning model training.
From Apple machine learning research
Research2 min read
Google DeepMind announced Gemini 4 Argon, a new AI model with a 1 million token limit, designed for complex workflows in coding, finance, legal, and cybersecurity.
From Google DeepMind blog
AI2 min read
NVIDIA, MIT, and Oxford developed Physis-Lang, a framework that adds physics reasoning to video captions. It improves physics accuracy in models.
From MarkTechPost
AI2 min read
H Company launched Holo4, a family of open-weight models for AI agents. They work on desktop, web, Android, and APIs with different sizes and licenses.
From MarkTechPost
AI1 min read
Salesforce Koa is built on Nvidia's open-weight Nemotron model and is trained to do sales, marketing, and customer-support tasks.
From TechCrunch AI
AI1 min read
Jensen Huang engaged in a conversation with President Trump at the All-In conference, discussing AI safety while demonstrating a new, unreleased foldable phone from Apple. This event highlighted Nvidia’s reliance on AI labs and Apple’s renewed interest in its hardware.
From TechCrunch AI
AI1 min read
Founders must consider how their AI products maintain value amidst rapid advancements from platform models like OpenAI and Anthropic. The session explores strategies for building defensible AI companies beyond simply competing with model capabilities.
From TechCrunch AI
Models1 min read
Fyxer leverages OpenAI’s models, combined with fine-tuning, memory, and user feedback, to automate inbox organization and email drafting. This approach aims to create an AI assistant that aligns with individual user communication styles.
From OpenAI news
Research1 min read
The BlueLM-GUI system, a 35B model, achieves strong performance on mobile GUI tasks through a real-device training approach. This flywheel system utilizes a multi-stage process of data salvage, real-device rollout, and evolving benchmarks to improve agent capabilities and transfer directly to production environments.
From arXiv cs.AI
Research1 min read
This paper introduces Affective Agent, a three-layer architecture for personalized intervention reasoning on wearable devices. The system uses a compact language model and physiological data to make intervention decisions without cloud dependency, improving closed-loop systems.
From arXiv cs.AI
Research1 min read
T-GADE, a system using thermodynamic genetic algorithms and LLMs, evolved structured artifacts like code paired with descriptions. Training experiments on the online bin-packing task showed a 29% reduction in median training excess at a temperature of 0.003, validating the approach’s effectiveness.
From arXiv cs.AI
Research1 min read
Researchers released Occamy-1.0, a 35 billion parameter model trained on a Qwen3.6-35B checkpoint. Evaluations across co-work benchmarks show it consistently performs among the strongest models of its size, achieving a competitive position on a cost-performance Pareto frontier.
From arXiv cs.AI
Research1 min read
Researchers introduced GT Bench, a benchmark with over 100,000 examples across four graph representations, to evaluate LLM performance on algorithmic graph problems. The Graph Theory Agent (GTA) system, combining a representation selector and a plan-and-decompose agent, achieved improved results on the GT Bench benchmark.
From arXiv cs.AI
LLMs1 min read
The research introduces Chopthin-Consensus Power Sampling (CCPS), a method for LLM decoding that preserves reasoning diversity and improves accuracy. Evaluation across open-weight models and benchmarks shows CCPS matches or exceeds the Power-SMC baseline in 14 of 15 settings, achieving gains of up to 10.6 percentage points.
From arXiv cs.CL
LLMs1 min read
The ORQA framework provides a method for testing large language model knowledge across 116 occupations using data sourced from trusted occupation-specific websites. Testing of 15 models revealed performance variations, with Claude Opus and GPT-5.4 achieving approximately 58-62% accuracy, while open-weight models showed 33-41% accuracy.
From arXiv cs.CL
Posts are drafted from public feeds by models OpenSmartRoute routes to - the same router, skill and metering customers use - and always link to the original source. Corrections: support.
Archive (100)
