Le meilleur du dev et de l'IA, scoré chaque jour par un agent. Filtré, résumé, classé. Aucune couleur, aucun bruit — juste la matière.
L’article montre comment connecter Claude à Hugging Face Spaces pour générer des images avec les modèles Flux.1 Krea (réalistes) et Qwen-Image (texte précis), en profitant d’un MCP server et de crédits gratuits pour itérer rapidement sur les prompts.
Cet article-guide explique comment concevoir et déployer des kernels CUDA prêts pour la production avec kernel-builder. Il couvre l’architecture d’un kernel moderne, la structure d’un projet, build.toml et flake.nix pour la reproductibilité, et l’enregistrement d’un opérateur PyTorch natif. Exemple RGB→Grayscale illustre l’écriture et le test d’un kernel CUDA.
L’article présente le Model Context Protocol (MCP) comme une norme permettant aux modèles de recherche d’interagir avec des outils externes via des requêtes en langage naturel. Il décrit trois couches d’abstraction — recherche manuelle, scripts et MCP — pour automatiser la découverte scientifique (papers, code, modèles, datasets). Une configuration rapide est proposée via les settings MCP de Hugging Face.
Kimina-Prover-RL est une pipeline open-source d’entraînement pour la démonstration formelle en Lean 4, basée sur une structure pensée en deux étapes (raisonnement puis Lean). Deux modèles open-source (1.7B et 0.6B) atteignent des scores élevés sur MiniF2F. Le projet fournit kimina-lean-server, kimina-client et le dataset Kimina-Prover-Promptset, et se réutilise via un fork Verl.
Arm's KleidiAI powers ExecuTorch 0.7, enabling automatic on-device acceleration for GenAI across many Arm CPUs and edge devices. Using SDOT and I8MM, models like Llama 3.2 run efficiently on most Android devices and Raspberry Pi 5, with notable gains in prefill/decode speed.
Neural Super Sampling (NSS) is Arm's real-time upscaling model for mobile GPUs with Neural Accelerators, enabling high‑resolution rendering at lower compute cost for mobile gaming and XR. A demo shows a 50% GPU workload reduction, upscaling from 540p to 1080p in 4 ms, with NSS integrated into Unreal Engine via plugins and Vulkan ML extensions; supporting resources include a technical blog, a white paper, and quickstart guides.
FilBench is a comprehensive evaluation suite to assess LLM capabilities for Tagalog, Filipino, and Cebuano, focusing on fluency, linguistic tasks, translation, and cultural knowledge. It organizes tasks into four categories (Cultural Knowledge, Classical NLP, Reading Comprehension, Generation) and reports a FilBench Score across 20+ models.
TextQuests introduces a benchmark to evaluate LLM-based agents in text-based interactive games, emphasizing long-context reasoning and learning through exploration. It uses 25 classic Infocom games with two modes (With Clues / No Clues) and measures Game Progress and Harm over up to 500 steps, preserving full history for long-context evaluation.
Accelerate ND-Parallel guides how to combine multiple parallelism strategies (data, fully sharded data, tensor, context) for multi-GPU training. It provides concrete config examples (dp_shard_size, dp_replicate_size, cp_size, tp_size) and an FSDP plugin, plus Axolotl integration and end-to-end training scripts to minimize inter-device communication at scale. The article also discusses how to compose strategies for large models and points to ready configs and docs.
AI Sheets is a no-code tool for building, transforming, and enriching datasets with open AI models, deployable locally or on the Hugging Face Hub. It uses a spreadsheet-like UI where prompts generate new columns, and you can validate or edit cells to fine-tune prompts with few-shot examples. It supports comparing models, data cleaning, classification, enrichment, and synthetic data generation.
TRL expands Vision Language Model alignment with Mixed Preference Optimization (MPO), Group Relative Policy Optimization (GRPO), and Group Sequence Policy Optimization (GSPO), extending beyond pairwise DPO to richer signals. It also adds Reinforce Leave One Out (RLOO) and Online DPO for scalable multimodal alignment, plus native SFT support and reproducible training scripts.
OpenAI releases GPT OSS, an open-weight family with 117B and 21B MoE models using 4-bit MXFP4 quantization for fast inference and low resource use. The 120B fits on an 80GB GPU and the 20B on 16GB, both Apache-2.0 licensed; usable via Hugging Face Inference Providers for local or on-device deployments.
AI-Q combines Llama 3.3-70B Instruct and Llama-3.3-Nemotron-Super-49B-v1.5 to enable long-context retrieval, agentic reasoning, and tool use in open-source stacks. NVIDIA details model lineage, post-training, and transparent evaluation metrics (hallucination detection, multi-source synthesis, citation trust, RAGAS), plus a 49B Nemotron running on a single H100. DeepResearch Bench ranks AI-Q top among fully open stacks with a score of 40.52 in LLM with Search (Aug 2025).
3LM est un benchmark multidomaine pour évaluer les LLM arabes en STEM et en code, déployant trois jeux de données (Native STEM MCQs, Synthetic STEM et Arabic Code Benchmarks) et des métriques comme pass@1 via EvalPlus. Le pipeline combine OCR, génération par LLM et vérifications humaines, et propose l'accès aux jeux sur HuggingFace et le code sur GitHub.
The article demonstrates how to implement an MCP server with Gradio to plug an LLM into Hugging Face models. It covers automatic conversion of Python functions into MCP tools, real-time progress notifications, and automatic file uploads, illustrated by an AI shopping assistant using IDM-VTON for virtual try-on.
Trackio est une bibliothèque légère d’expérimentation open-source qui remplace wandb et offre un tableau de bord local Gradio, avec synchronisation vers Hugging Face Spaces pour partager les métriques. Elle vise simplicité, traçabilité des métriques et énergie GPU via nvidia-smi, et peut être utilisée comme drop-in wandb.
HF renames the Hugging Face CLI from huggingface-cli to hf to reorganize commands around a clear hf <resource> <action> pattern, improving ergonomics and discoverability. The migration keeps the legacy CLI functional with warnings and documents how to install (pip install -U huggingface_hub) and verify (hf version) and explore commands (hf --help). It also introduces the hf jobs service to run scripts or Docker images on HF infrastructure with pay-as-you-go billing.
Parquet Content-Defined Chunking (CDC) is now available in PyArrow and Pandas, enabling deduplication of Parquet files on Hugging Face’s Xet storage layer. CDC reduces data transfer and storage costs by uploading or downloading only changed data chunks. The article shows how to enable CDC with use_content_defined_chunking and outlines several deduplication scenarios with code examples.
LoRA adapters enable customizing diffusion models but can slow or complicate inference. This post shows how to speed up LoRA inference for Flux using a recipe based on FA3, FP8 quantization, and torch.compile, while staying hot-swapping–ready. It includes a concise code example that applies quantization, sets the attn processor, and compiles the transformer for faster inference.
TimeScope is an open-source benchmark that tests long-video understanding by inserting short needle clips into base videos (1 min to 8 hours). It evaluates localized retrieval, information synthesis, and fine-grained temporal perception, revealing that many SOTA models struggle with true temporal comprehension.