DeepSeek V4.1-Flash-1017: Open Weights, Beats Opus 5 on Code
DeepSeek released V4.1-Flash-1017 today under MIT — 77.0 on DeepSWE beats Claude Opus 5, open FP8 weights on Hugging Face, and a $0.75/1k ...
Liquid AI LFM2.5-24B-A2B: Run an Agent Model on Your Laptop
Liquid AI released LFM2.5-24B-A2B, a 24B open-weight MoE model that runs at 184 tok/s on ...
AI Agents Are Writing CUDA Kernels That Beat PyTorch
AI agents now generate CUDA kernels with 2.6x speedups over PyTorch Eager. We cover CUDA ...
Burn 0.22.0: Generic-Free Tensors, LoRA/QLoRA, 15x Faster
Burn 0.22.0 removes backend generics from all models, cuts rebuild times by up to 15x, ...
EmbeddingGemma 2 Runs 5-Modality Search in 567MB RAM
Google DeepMind shipped EmbeddingGemma 2: a 740M open-weight model mapping text, code, images, video, and ...
PyTorch 2.14: Fault-Tolerant Training and Apple Silicon Fixes
PyTorch 2.14 ships fault-tolerant distributed training, fixes an 8.5x Apple Silicon decode bug, introduces NVGEMM epilogue fusion, and stabilizes TorchVision ABI.
Kubeflow Is Now CNCF Graduated — What ML Engineers Must Do
Kubeflow is now CNCF Graduated. Here is what the unified TrainJob API, enterprise adoption numbers, ...
IBM and NASA Open-Source Prithvi-Lunar: Use It Now
IBM and NASA released Prithvi-Lunar today — an Apache-2.0 foundation model for the Moon trained ...
humain-m3 Is in Preview: What Arabic NLP Developers Need to Know
HUMAIN launched humain-m3 at LEAP 2026 — a 428B MoE Arabic LLM built on MiniMax-M3, ...
Amazon Mechanical Turk Shuts Down Sept. 30: Act Now
Amazon Mechanical Turk closes September 30, 2026. Here is what breaks in SageMaker Ground Truth ...