Samsung zHBM: Memory Stacked on AI Chips — 8x HBM5 Speed
Samsung unveiled zHBM at FMS 2026 — a memory architecture that stacks HBM directly above AI accelerators for 8x HBM5 bandwidth. Here is ...
Microsoft Orchard: 73% SWE-bench With 3B Parameters
Microsoft Research open-sourced Orchard, a Kubernetes-native agent training framework that hits 73% on SWE-bench Verified ...
Thinking Machines Inkling: 975B Open-Weight Model for Fine-Tuning
Thinking Machines released Inkling — a 975B MoE open-weight model. Here is the architecture, benchmark ...
GLM-5.2 Beats GPT-5.5 on SWE-bench — And You Can Self-Host It
Z.ai’s GLM-5.2 outscores GPT-5.5 on SWE-bench Pro with MIT-licensed open weights at 6x lower cost. ...
IBM CodeAlchemy: ~1 Trillion Tokens of Open Code Data
IBM Research open-sourced CodeAlchemy: 976.6B-token synthetic code dataset, 15 languages, execution traces. Apache 2.0. On ...
Kubeflow 1.11: MLOps Gets a pip install Moment
Kubeflow 1.11 ships a Python-first SDK, YAML-free LLM fine-tuning for Llama 3.2, and a new OptimizerClient. Here is what changed and what to ...
NVIDIA Nemotron-Labs TwoTower: 2.42x Faster Inference, No Retraining Required
NVIDIA Nemotron-Labs TwoTower retrofits a pretrained 30B AR model into a diffusion decoder running 2.42x ...
FLUX 3: Black Forest Labs Launches One Model for Video, Image, and Robots
Black Forest Labs launched FLUX 3 on July 23 — one model generating video with ...
Google TabFM Beats Tuned XGBoost. Here Is When That Actually Matters.
Google TabFM does zero-shot tabular classification and regression with no training. It beats Optuna-tuned XGBoost ...
PyTorch 2.13: FlexAttention on Apple Silicon Is 12x Faster
PyTorch 2.13 brings FlexAttention to Apple Silicon with hand-written Metal kernels, delivering up to 12x ...