ML Frameworks & Training

LaunchPrimary topic: ML Frameworks & TrainingAlso: Computer Vision

Bring multimodal semantic search to the edge with EmbeddingGemma 2

Crest summary (AI-written):Google released EmbeddingGemma 2, a 740M open-weight multimodal model that maps text, images and video for semantic search on edge devices.

Companies: Google

LaunchPrimary topic: Embedded, IoT & HardwareAlso: ML Frameworks & Training

Native Model Authoring for Edge with LiteRT Tensor API

Crest summary (AI-written):Google introduced the LiteRT Tensor API for authoring models natively for edge devices.

Companies: Google

LaunchPrimary topic: LLM InfrastructureAlso: ML Frameworks & TrainingAlso: Cloud Infrastructure

Announcing New vLLM and SGLang Inference Engines on TPU

Crest summary (AI-written):Google announced native PyTorch backends for the vLLM and SGLang inference engines on TPU, powered by TorchTPU.

Companies: Google

EngineeringPrimary topic: SecurityAlso: ML Frameworks & Training

Securing our frontier model training environment

Crest summary (AI-written):OpenAI published an account of how it secures the environment used to train its frontier models.

Companies: OpenAI

UpdateTopic: ML Frameworks & Training

Integrating TorchTPU into TorchTitan and Optimizing Models for TPU

Crest summary (AI-written):Google integrated TorchTPU, a native torch.device("tpu") backend, into PyTorch, letting developers train models on TPUs.

Companies: AWS · Google

EngineeringPrimary topic: ML Frameworks & TrainingAlso: Observability

Profiling TorchNeuron (Part 1): System-level profiling on AWS Trainium.

Crest summary (AI-written):Hugging Face published the first part of a series on profiling TorchNeuron, covering system-level profiling on AWS Trainium.

ResearchTopic: ML Frameworks & Training

Decoding cosmic signals with deep learning and Keras

Crest summary (AI-written):Google Developers Blog published a post on using deep learning and Keras to decode cosmic signals in astroparticle physics.

Companies: Google

LaunchPrimary topic: LLM InfrastructureAlso: ML Frameworks & Training

Autonomous LLM post-training with Tunix on TPUs

Crest summary (AI-written):Google's autofinetune project introduces an autonomous research loop that fully automates LLM post-training workflows, including fine-tuning, on TPUs via Tunix.

Companies: Google

EngineeringPrimary topic: ML Frameworks & TrainingAlso: Cloud Infrastructure

Reproducing Olmo 3 7B Pre-training in MaxText: case study of large scale training on TPUs

Crest summary (AI-written):Google's MaxText team reproduced Ai2's Olmo 3 7B language model from scratch on Google Cloud TPUs using JAX/XLA, matching the original training run.

Companies: Ai2 · Google

EngineeringPrimary topic: Generative Image & VideoAlso: ML Frameworks & Training

Accelerating Spatio-Temporal Attention for Video Diffusion on TPUs

Crest summary (AI-written):Google Developers Blog describes implementing Sparse VideoGen to address the quadratic latency bottleneck of self-attention in high-resolution video diffusion models on TPUs.

Companies: Google

LaunchPrimary topic: LLM InfrastructureAlso: ML Frameworks & Training

Introducing Clef: our open-source decision models, and new RL fine-tuning platform

Crest summary (AI-written):Cloudflare introduced Clef and Clef-flash, open-source decision models hosted on Workers AI for classification and agentic workflows, plus a reinforcement learning platform for fine-tuning decision models with developers' own data.

Companies: Cloudflare · Hugging Face

LeadershipTopic: ML Frameworks & Training

Jun Kim, oMLX creator and maintainer, joins Hugging Face to support the MLX community

Crest summary (AI-written):Jun Kim, creator and maintainer of oMLX, has joined Hugging Face to support the MLX community.

Companies: Hugging Face

LaunchPrimary topic: NLP & SpeechAlso: ML Frameworks & Training

tokenizers v1: encode, decode and scaling, measured

Crest summary (AI-written):Hugging Face released tokenizers v1, a major update to its text tokenization library, with encode and decode performance and scaling measured in the announcement.

Companies: Hugging Face

EngineeringPrimary topic: ML Frameworks & TrainingAlso: LLM Infrastructure

Async GRPO with LoRA across HF Jobs: a bucket, a proxy, and no NCCL

Crest summary (AI-written):Hugging Face published a method for running asynchronous GRPO reinforcement learning with LoRA adapters across its Jobs infrastructure, using a bucket and a proxy instead of NCCL.

Companies: Hugging Face

ResearchTopic: ML Frameworks & Training

Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps

Crest summary (AI-written):Hugging Face published a method for fine-tuning a 350M-parameter model to improve structured outputs using 100 GRPO steps.

Companies: Hugging Face

EngineeringTopic: ML Frameworks & Training

Training a coding model to paint watercolours with TRL and OpenEnv

Crest summary (AI-written):Hugging Face published a walkthrough on training a coding model to produce watercolour-style output using its TRL library and OpenEnv environment.

Companies: Hugging Face

LaunchPrimary topic: Browsers & ExtensionsAlso: ML Frameworks & Training

Introducing @huggingface/kernels: 200+ WebGPU Kernels for Local AI

Crest summary (AI-written):Hugging Face released @huggingface/kernels, a library of more than 200 WebGPU kernels for running AI models locally in the browser.

Companies: Hugging Face

EngineeringTopic: ML Frameworks & Training

Training and Finetuning Multi-Vector Embedding Models with Sentence Transformers

Crest summary (AI-written):Hugging Face published guidance on training and finetuning multi-vector embedding models using Sentence Transformers.

Companies: Hugging Face

OtherPrimary topic: RAGAlso: ML Frameworks & Training

Multi-Vector (Late Interaction) Embedding Models with Sentence Transformers

Crest summary (AI-written):Hugging Face published guidance on using multi-vector, late-interaction embedding models through the Sentence Transformers library.

Companies: Hugging Face

ResearchPrimary topic: ML Frameworks & TrainingAlso: LLM Infrastructure

Making Knowledge Distillation Cheap Enough to Run at Scale

Companies: Hugging Face

UpdatePrimary topic: Generative Image & VideoAlso: ML Frameworks & Training

Bringing Nunchaku 4-bit Diffusion Inference to Diffusers

Crest summary (AI-written):Hugging Face announced that Nunchaku 4-bit diffusion inference is now available in Diffusers. The integration lets users run diffusion models at 4-bit precision through the Diffusers library.

Companies: Hugging Face

EngineeringTopic: ML Frameworks & Training

Profiling in PyTorch (Part 3): Attention is all you profile

Crest summary (AI-written):Hugging Face published the third part of its PyTorch profiling series, focused on attention mechanisms.

Companies: Hugging Face

LaunchPrimary topic: LLM InfrastructureAlso: ML Frameworks & Training

Native-speed vLLM transformers modeling backend

Crest summary (AI-written):Hugging Face introduced a native-speed vLLM transformers modeling backend on 2026-07-08, letting transformer models run through vLLM without a separate conversion step.

Companies: Hugging Face