The best of dev and AI, scored every day by an agent. Filtered, summarized, ranked. No color, no noise — just the substance.
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.
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.
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.
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.
OpenAI a publié Deep Research, un système qui navigue sur le Web pour résumer le contenu et répondre par le résumé. L’article présente une reproduction open-source du cadre agentique (CodeAgent) et les résultats sur le benchmark GAIA (≈67% en 1-shot, 47,6% niveau 3). Les auteurs visent une open-source du cadre et les prochaines étapes de reproductibilité.
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.
Introducing DABstep, a benchmark of 450+ real-world data analysis tasks to evaluate multi-step reasoning in AI agents. The study finds current top agents reach only about 16% accuracy, underscoring a large gap to reliably tackle real data tasks that mix structured data and unstructured documents.
Open-R1: Update #1 résume les progrès pour répliquer le pipeline d’entraînement et les données synthétiques de DeepSeek-R1 (MATH-500, GRPO dans TRL 0.14, DeepSpeed ZeRO, vLLM). Le post évoque aussi les défis de longueur des générations et propose des ressources communautaires ainsi qu’un leaderboard public.
Replaying the DeepSeek-R1 'aha moment', this post uses Group Relative Policy Optimization (GRPO) and the Countdown Game to train an open model via RL. It details a distributed setup with DeepSpeed and vLLM on 4× NVIDIA H100 GPUs and explains how GRPO replaces a value function with group-based baselines. The aim is self-verification and search abilities learned with minimal human data, illustrating a concrete RL workflow for LLMs.
2024 saw major open-source breakthroughs in AI art, with a shift to diffusion transformers (DiT) and flow matching in text-to-image generation. Flux.1 achieved state-of-the-art results, outperforming some closed models, as open releases like Stable Diffusion 3, Stable Diffusion 3.5, AuraFlow and HunyuanDiT expanded the open ecosystem. The article also highlights personalization advances via SDXL and hints at 2025’s ongoing open-source momentum.
The article shows how to deploy and fine-tune DeepSeek R1 models on AWS with Hugging Face. It covers Inference Endpoints, Bedrock, SageMaker, and EC2 Neuron deployments, plus notes on pricing and upcoming Inferentia support.
Open-R1 aims to reproduce DeepSeek-R1's reasoning capabilities using reinforcement learning with minimal human supervision, and to reveal training data and hyperparameters. It builds on DeepSeek-V3, a 671B Mixture-of-Experts model, and contrasts the RL-only DeepSeek-R1-Zero with the refined DeepSeek-R1.
Hugging Face déploie quatre fournisseurs d’inférence serverless (fal, Replicate, Sambanova, Together AI) directement sur les pages des modèles du Hub, et les intègre dans les SDK JS et Python. Les utilisateurs peuvent configurer leurs clés API ou opter pour un routage via Hugging Face, avec des exemples d’utilisation via InferenceClient pour appeler des modèles comme DeepSeek-R1.
CocoIndex fournit une couche incrémentale qui maintient des contextes AI frais en réutilisant uniquement le delta entre sources variées, avec des runtimes Python/Rust et des exposé
Documente un générateur de romans IA, avec architecture modulaire, configuration détaillée, et un flux de travail pas-à-pas (setup, génération et vérifications).
Skill Claude prêt à l'emploi pour construire, tester et automatiser iOS avec 22 scripts: Xcode via xcodebuild, navigation via idb/simctl, et sorties résumées.