Pre A

Black Forest Labs releases FLUX 3 Image Klein, a 5.6B open-weights model running on 8 GB VRAM: Black Forest Labs released FLUX 3 Image Klein, a 5.6B open-weights image model with native 4K generation, pixel-precise editing, and up to ten reference images. It runs on GPUs with 8 GB VRAM, has 38% fewer parameters than its predecessor, and beats proprietary competitors in internal tests.

AI Models • The Decoder

arXiv paper presents ARIA, a framework that learns to repair LLM multi-agent executions at runtime: An arXiv paper presents Agent Runtime Inspection and Amendment (ARIA), a framework for learning when and how to repair ongoing LLM-based multi-agent executions. It converts execution-verified repairs into reusable supervision, enabling a learned model to intervene at runtime rather than repeatedly testing alternatives.

AI Research • arXiv

Microsoft unveils MAI-Base-2 MoE with 2.8T total, 80B active parameters: Microsoft unveiled MAI-Base-2, a mixture-of-experts model with 2.8T total and 80B active parameters. It claims 4x-12x more pretraining FLOPs-efficiency than DeepSeek-V4-Pro-Base. Microsoft also cited MAI-Thinking-1, a general-purpose foundation model post-trained on MAI-Base-1 announced in June.

AI Models • Techmeme

ACED-BENCH tests whether LLM agents acquire and act on causal evidence: ACED-BENCH is a neuro-symbolic benchmark for evaluating whether LLM agents acquire and act on causal evidence. It separates a readable scientific surface from a formal causal world and preserves objective ground truth. Agents choose experiments, interpret evidence, and make causal decisions.

AI Benchmark • arXiv

StepFun releases open weights for Step 5 Preview, a 600B model with 27B active parameters: StepFun released open weights for its Step 5 Preview model, with 600B total and 27B active parameters and a 1M context window. It published a technical blog describing the model's 92-layer design, sparse grouped-query attention, and long-horizon agent training.

StepFun • Pandaily

LHTB 2.0 expands its benchmark to 77 tasks across 13 domains: LHTB 2.0 extends Long-Horizon-Terminal-Bench, preserving 46 original tasks while correcting task and evaluation-infrastructure defects. It adds 31 extended tasks, giving 77 tasks across 13 domains and eight task types. Extended tasks are produced through a systematic reconstruction-task pipeline.

AI Agent Benchmark • arXiv

Nature Will unveils Natura-1, a 50B MoE pretrained on 1M hours of human video: Nature Will's Natura-1 is a 50-billion-parameter MoE pretrained on one million hours of first-person human video. The model was then adapted with limited robot data to control 20-DoF robot hands.

Robotics • Pandaily

Google Just Caught Up (Gemini-4 Argon): Matt Wolfe posted a hands-on test of Gemini 4 Argon that included coding games and a 3D-printing Gary Busey task. He compares it with Claude and ChatGPT and notes which tasks impressed him and which got stuck.

Artificial Intelligence • Matt Wolfe

Amex introduces framework to prevent unauthorized AI agent purchases: American Express introduced a framework to ensure AI agents only complete purchases their users have authorized. The initiative aims to prevent unauthorized purchases and related disputes as AI-powered shopping grows. Specific availability and enforcement details were not disclosed.

American Express • TechCrunch AI

AGIBOT • AGIBOT

Chinese AI robot maker AGIBOT nears new funding round at about RMB 50 billion valuation: AGIBOT is nearing completion of a new funding round of approximately RMB 1 billion to RMB 1.5 billion, valuing the Chinese AI robot maker at about RMB 50 billion, according to investors cited by Caijing. Founded in 2023, AGIBOT began preparing for a Hong Kong listing in July.