Le meilleur du dev et de l'IA, scoré chaque jour par un agent. Filtré, résumé, classé. Aucune couleur, aucun bruit — juste la matière.
L2D is introduced as the world's largest multimodal self-driving dataset (90+ TB, 5000+ hours, 30 cities). It provides expert and student driving policies with natural language instructions and future waypoints, enabling end-to-end training of automotive AI and inviting community contributions.
This tutorial shows how to run LLMs on-device via a React Native app, downloading GGUF models from Hugging Face and loading them with llama.rn (llama.cpp). It covers model sizing and quantization formats (Q4_0, Q3_K, IQ2_XXS, etc.), and outlines a complete app workflow from environment setup to debugging, all aimed at offline, privacy-preserving mobile AI.
Cohere's Aya Vision unveils 8B and 32B multilingual vision-language models for 23 languages, with strong results on AyaVisionBench and mWildVision. It uses synthetic annotations, multilingual data expansion, and model merging, and handles high-res images via dynamic tiling and Pixel Shuffle downsampling before the vision-language connector and LLM. Training is in two stages: vision-language alignment with frozen encoders, then supervised fine-tuning across multilingual tasks.
Hugging Face and JFrog announce a security partnership to integrate JFrog's scanner into the Hugging Face Hub, reducing false positives by analyzing code inside model weights rather than relying on pattern matching. The collaboration targets risks from serialization formats like pickle that can enable code execution, and all public repos will be scanned automatically as you push.
Cet article présente Arize Phoenix pour tracer, évaluer et déboguer les agents IA en temps réel. Il démontre comment activer le tracing via OpenTelemetry et OpenInference, et propose des étapes concrètes pour installer smolagents, configurer un agent et lancer des tests. Des extraits de code et d’utilisation servent de guide pratique pour améliorer la traçabilité et la performance.
Hugging Face and IISc/ARTPARK partner to enable access to Vaani, India's diverse open-source, multi-modal dataset. Vaani targets 54 languages, 150k hours of speech and 15k hours of transcribed data from about 1 million people across 773 districts, with Phase 1 already open-sourced and Phase 2 expanding reach. This dataset supports ASR, TTS, language identification, and speaker verification, empowering robust Indic-language models and code-switching applications.
FastRTC is a Python library that simplifies building real-time audio and video AI apps. It offers automatic voice detection and turn-taking, a built-in Gradio UI, WebRTC/WebSocket support, and easy deployment to FastAPI, plus utilities for STT/TTS and phone-based access. The post walks through a hello-world echo example and a level-up LLM chat flow with SambaNova/OpenAI integration.
Cet article présente le déport du décodeur VAE vers une endpoint distante pour réduire la mémoire et la latence lors de l’inférence de modèles de diffusion en latent. Il détaille l’intégration avec diffusers et remote_decode et montre des cas d’usage (Stable Diffusion v1.5, Flux, HunyuanVideo) via des exemples de code et les avantages (aucun stockage des données, open source).
SigLIP 2 expands Google's multilingual vision-language encoder family by adding additional training objectives to SigLIP's sigmoid loss, boosting semantic understanding, localization, and dense features. It outperforms SigLIP across scales on zero-shot classification, image-text retrieval, and transfer tasks, and introduces a dynamic resolution (naflex) variant for aspect-ratio-sensitive downstream work. The release catalogs multiple models (Base, Large, So400m, Giant) with varied patch sizes, 2
SmolVLM2 rend la compréhension vidéo accessible sur tous les devices via trois tailles (2.2B, 500M et 256M) et des APIs MLX dès le départ. Par rapport à la génération précédente, il optimise la mémoire et excelle sur Video-MME, avec des démos et une interface interactive pour tester la vision et la compréhension vidéo même dans un Colab gratuit.
Google publie PaliGemma 2 Mix, une famille de modèles vision-langage fine-tunés sur OCR, captioning et autres tâches. Disponibles en 3B/10B/28B et résolutions jusqu'à 896x896, ils permettent d’estimer les performances après fine-tuning sur des tâches en aval. L'article détaille des prompts open-ended et des prefixes (caption, describe, ocr, answer) ainsi que des invites pour détection et segmentation, avec une démo.
Hugging Face adds Hyperbolic, Nebius AI Studio, and Novita as serverless inference providers, integrated in the Hub and accessible via JS/Python SDKs. It outlines two usage modes (custom provider keys or routing through HF) with code examples for DeepSeek-R1 and Flux.1, and explains billing depending on the routing mode.
Math-Verify a remis à plat l'évaluation des LLM sur le Open LLM Leaderboard, en réévaluant 3 751 modèles sur 1 324 problèmes de maths difficiles. L'article explique les failles de l'ancienne méthode (format de réponse, parsing SymPy) et décrit les améliorations qui permettent une comparaison plus juste et robuste des modèles.
Fireworks.ai est désormais un fournisseur d’inférence pris en charge sur Hugging Face Hub. L’article présente comment lancer de l’inférence serverless via Fireworks.ai sur Python, JS et HTTP, avec des exemples sur des modèles comme DeepSeek-R1 et Llama-3, et détaille la facturation et les crédits PRO.
The piece breaks down how to cost-effectively run 1B+ classifications or embeddings at scale, analyzing model architectures, hardware options, and deployment choices. It offers a framework to estimate cost and latency, plus a practical stack (Inference Endpoints, Hugging Face Hub, Infinity, k6) to benchmark and optimize throughput.
The post describes tooling to build video-generation datasets, extending the img2dataset approach to videos. It presents a three-stage pipeline (Acquisition, Pre-processing/Filtering, Processing) using yt-dlp, Video to Scenes, watermark and aesthetic/NSFW checks, motion scoring, and Florence-2-based captions/OCR to filter data for fine-tuning.
The article argues that content-defined chunking (CDC) is only a means to speed up data movement, not the ultimate goal of deduplication. To scale, the team moves from per-chunk transfers to aggregation: blocks up to 64MB reduce CAS entries by ~1000x, while shards map files to blocks and detect changes, enabling faster uploads/downloads.
OpenR1-Math-220k is a large-scale math-reasoning dataset built on 512 H100s, with two (often four) solutions per problem and about 800k reasoning traces. It uses automated filtering (Math Verify) and LLama3.3-70B-Instruct as a judge, and achieves high throughput with vLLM and SGLang (~300k solutions/day). The update also details distillation results and a pipeline extensible to other domains.
Open Arabic LLM Leaderboard evolved from fragmented benchmarks to a unified platform. Since May 2024, OALL hosts 14 benchmarks; later Balsam Index added ~1,400 datasets and 50,000 questions, AraGen launched 3C3H with private test cycles, and SEAL introduced a private Arabic leaderboard with human-preference evaluation. The ecosystem drew 46k visitors and 700+ models from 180+ organizations.
π0 and π0-FAST are Vision-Language-Action models for generalist robot control. They use flow-matching for real-time action trajectories (50 Hz) across seven platforms and 68 tasks, and introduce fast attention techniques (FlashAttention2, FlexAttention) to handle 2D masks and cross-embodiment training. The post also points to Hugging Face LeRobot repos for code and pretrained models.