AI BriefingHuggingFace Blog news
The latest AI and developer news from HuggingFace Blog, with short summaries.
Visit website →PickNVIDIA Releases Nemotron 3.5 Embed, Rerank, and Parse Models for Agent Retrieval
NVIDIA releases five Nemotron 3.5 models for embedding, reranking, and parsing, with the 8B Embed model topping ViDoRe v3.1 and the suite reducing agent token costs by up to 14%.
PickNVIDIA Introduces Cosmos3-H-Dreams: Real-Time Generative Surgical Simulator Trained on Clinical Cholecystectomy Data
The system is the first real-time generative surgical simulator trained on clinical cholecystectomy recordings to support live kinematic control.
Ai2 releases FairStress benchmark revealing asymmetric robustness in LLM decision-making
The 13-million-case benchmark tests 15 models and finds that existing debiasing techniques often fail to fix this identity-based vulnerability without sacrificing accuracy.
IBM Research ALTK-Evolve tops tau2-bench retail leaderboard with learned guidelines on GPT-4.1
The method uses GPT-4.1 to extract behavioral guidelines from agent trajectories, improving reliability and accuracy without weight updates or human annotation.
Hugging Face integrates TorchTPU for native PyTorch support on TPUs
Transformers, Diffusers, and Kernels now run natively on TPUs using the TorchTPU backend, eliminating the need for torch_xla or JAX.
Hugging Face releases AnyLanguageModel 1.0 for local and hosted LLM inference on Apple platforms
The 1.0 release adds support for OpenAI, Anthropic, and Gemini APIs alongside local models like Gemma 4 E2B and Qwen3.8 27B.
PyTorch 2.15 Adds FP8 Support and Boosts LLM Decoding Speed on Apple Silicon
The update introduces native FP8 support and optimizes Metal kernels, reducing driver memory peaks from 270.135 GiB to 8.108 GiB for video decoding.
Hugging Face Integrates TorchNeuron for Native PyTorch Support on AWS Trainium
The new backend enables eager execution and native distributed APIs in Transformers, Diffusers, and Kernels, replacing the previous Optimum Neuron workflow.
DeepSeek Releases V4.1-Flash-1017
Terminal-Bench scores rose from 31.2 to 42.3, with improved performance on agent and vision benchmarks.
NVIDIA and AI4Bharat Release 2.8 Trillion Token Indic Pretraining Dataset and Open Data Generation Pipeline
The dataset covers 22 Indic languages with native and romanized scripts, generated via an open-source NeMo Data Designer pipeline.
Reflection releases Beam: a 501B MoE open model with 1M context and specialized agentic capabilities
The Apache 2.0-licensed model features 23B active parameters, a 1M token context window, and day-0 support for vLLM, SGLang, and Transformers.
PickIBM Releases Granite Switch 4.1 Models Combining Chat, Decisions, and Speech on Single Endpoint
The new 8B and 30B models integrate a Decision aLoRA, Guardian aLoRA, and the fastest ASR model on the OpenASR Leaderboard while maintaining base-model VRAM usage.
Swiss AI Initiative Releases Apertus 1.5: Multimodal Open-Weight Models with 262k Context
The new 8B and 70B models add native image and audio understanding, thinking mode, and a 4x larger context window under an Apache 2.0 license.
Meta Releases Muse Spark Open: 816B-Parameter Open-Weight Multimodal MoE Model
The model features 42B active parameters per token, a 1M-token context window, and supports text, image, audio, and video inputs.
Black Forest Labs Releases FLUX.3 Image Klein with Layout Control and Multi-Reference Support
The new model supports text-to-image, multi-reference editing, and bounding-box layout control, with distilled checkpoints generating 1024x1024 images in ~2 seconds on H100 GPUs.
Ai2 releases HiRO-ACEv2 for extreme weather simulation at 3 km resolution
The new system generates global 3 km climate data at approximately 1/300th the cost of the X-SHiELD atmospheric model.
Liquid AI Releases LFM2.5-24B-A2B: A 24B-Parameter Reasoning Model with 2B Active Parameters for On-Device Agents
The open-weight model achieves 184 tokens/s on an Apple M5 Max and scores 48.09 on LiveClaw, outperforming larger models in agentic tasks.
Ai2 Replaces Priority Scheduler with GPU Time Budgets, Cutting Debug Wait Times from 2 Hours to 30 Seconds
The new system reduced p90 debug workload wait times from 2 hours to 30 seconds and cut on-call maintenance toil by 74% across clusters of thousands of NVIDIA GPUs.
Hugging Face Blog Details Six AI Models Built by ML Intern Agent for About USD 103
The ML Intern agent trained, evaluated, and published six specialized models, including a 0.8B prompt rewriter and a 4-step text-to-image distillation, on Hugging Face hardware for about USD 103.
Liquid AI Releases Open-Weight d1 Decision Models for Edge Inference
The new d1-3B model achieves a score of 48.57 on the Decision Index 0.2.1, outperforming larger models like Decider 35B-A3B.