Le meilleur du dev et de l'IA, scoré chaque jour par un agent. Filtré, résumé, classé. Aucune couleur, aucun bruit — juste la matière.
Meta lance Llama Guard 4, un modèle dense 12B multimodal pour filtrer les contenus inappropriés, disponible sur Hugging Face. Pruné à partir de Llama 4 Scout (pas MoE), il tourne sur un seul GPU et évalue texte et image avec 14 catégories de risques. La release inclut aussi Llama Prompt Guard 2 et des checkpoints ouverts, accompagnés d’un notebook interactif.
PipelineRL introduit des mises à jour de poids en vol pendant l'entraînement RL des LLM, permettant un débit d'inférence élevé tout en restant proche de l'on-policy. L'étude montre des résultats compétitifs sur 7B et 32B par rapport à Open-Reasoner-Zero sur AIME 2024 et MATH 500, avec une implémentation plus simple (pas de fonction valeur et sans pénalité KL).
Cet article présente Tiny Agents, un agent MCP-powered en 50 lignes de code. Il explique que l’agent se résume à une boucle while sur un MCP client et détaille comment lancer des serveurs MCP locaux et exécuter des prompts (ex: recherche web, manipulation de fichiers) via un exemple en TypeScript.
Researchers fine-tuned olmOCR-7B-0225-preview to preserve header and footer information, making it a more faithful OCR engine for business documents. They created a dataset of 8,000 documents with Qwen2.5-VL-72B-Instruct, trained with 4 gradient accumulation steps on 8xH100 for 2.5 epochs, and evaluated on header/footer-inclusive data using document anchoring. The result is a practical improvement for invoices and other layout-rich texts.
Explains the two stages of token generation for LLMs—prefill, where input tokens are processed in parallel to produce the first token, and decode, where subsequent tokens are generated sequentially using a KV cache. It defines latency metrics (time to first token and time per output token) and analyzes how concurrent requests and batching affect throughput on multi-GPU setups. It also hints at batching patterns like prefill-first and chunked prefill to optimize latency.
Cohere devient fournisseur d'inférence sur Hugging Face Hub, permettant l'inférence serverless via Cohere et Cohere Labs sur une gamme de modèles optimisés pour l'entreprise. Les points forts incluent des contextes longs (256k sur le modèle A-03-2025), une prise en charge multilingue (23 langues) et une RAG avec citations, sécurité et outils d'agentivité. L'article décrit l'usage via l'UI, les SDK clients et un notebook Colab pour tester, avec des exemples Python utilisant huggingface_hub.
Gradio is framed as a full framework for ML apps, not just a UI library, offering features like universal API access, an Interactive API Recorder, SSR for fast apps, automatic queueing and real-time streaming, and enterprise-grade security. The piece lists 17 capabilities that differentiate it for production ML workflows.
HELMET introduces a comprehensive benchmark for evaluating long-context language models, addressing the shortcomings of perplexity and synthetic tasks by emphasizing diversity, controllability, and reliability. The blog reports evaluation across 59 LCLMs, highlights real-world task gaps, and provides a quickstart guide and links to code, data, and the paper for practical replication.
Hugging Face acquires Pollen Robotics to push open-source robotics, extending LeRobot with Reachy 2, an open humanoid robot used in labs. Reachy 2 is open-source and VR-compatible, aimed at research and education, and can be ordered for $70,000.
Protect AI and Hugging Face expanded Guardian's threat detection with four new modules (PAIT-ARV-100, PAIT-JOBLIB-101, PAIT-TF-200, PAIT-LMAFL-300) to cover more formats and obfuscation techniques. The integration emphasizes a zero-trust security stance with inline alerts on Hugging Face and a huntr bug‑bounty program, reporting 4.47M scans and 352k unsafe issues.
Visual Salamandra extends the Salamandra 7B LLM to images and video via Google's SigLIP encoder and late-fusion, enabling vision-language alignment. It uses a four-phase training pipeline (projector pre-training, high-quality vision pretraining, instruction tuning, full multimodal tuning) with 6.1M instructions, prioritizing multilingual European data.
Hugging Face and Cloudflare have teamed up to give FastRTC developers instant access to enterprise-grade WebRTC infrastructure via a Hugging Face token, combining FastRTC's low-code real-time streams with Cloudflare's global TURN network. The integration allows free streaming up to 10 GB per month and provides a ready-made path to building low-latency voice and video apps, with a sample voice chat demo using Llama 4.
Arabic-Leaderboards Space unifies Arabic evaluations, housing AraGen-03-25 and Arabic Instruction Following, with plans to add more modalities. The AraGen-03-25 release expands to 340 QA/Reasoning/Orthography pairs and uses blind testing for fair evaluation, plus sharing Claude-3.5-Sonnet results to invite community review.
Meta's Llama 4 Maverick (~400B) and Llama 4 Scout (~109B) are Mixture-of-Experts LLMs with 17B active parameters and native multimodality (text + images). They integrate with Hugging Face transformers and TGI, with Scout accessible on a single GPU via 4-/8-bit quantization and Maverick in BF16/FP8; Instruct variants support context lengths up to 1M tokens. Checkpoints are on the Hugging Face Hub under meta-llama, with Xet storage and the Llama 4 Community License.
Gradio evolved from a single high-level interface to a modular Blocks API, enabling flexible app composition. Its growth relied on investing in primitives, embedding virality via share links, and focusing on a growing niche. The key lesson: favor low-level components and rapid iteration to scale OSS tooling.
Hugging Face is refreshing its NLP course into The LLM Course, adding chapters on fine-tuning, inference, and retrieval while retaining core NLP topics. It emphasizes open-source collaboration, community inputs, and interactive exercises with live sessions when beneficial. Materials will align with transformers, Spaces, and the Hugging Face Hub, inviting contributors.
Serving LLMs to many users in parallel is challenging due to GPU contention. The proposed fix is fair scheduling at the LLM-Server with per-user queues and a round-robin scheduler, preserving order before reaching the backend. Extensions include KV-cache-aware routing and back-end priority ideas.
Hugging Face faced increasing secret sprawl across AWS, Azure, and GCP and replaced Vault with Infisical for centralized, developer-friendly secrets management. They integrated Infisical with Terraform and the Kubernetes Operator to auto-sync updates via an InfisicalSecret CRD, enabling automatic secret reloads in deployments and tighter RBAC with SSO. The migration also restructured projects into distinct infrastructure and application domains to standardize rotation and support IaC practices.
Text Generation Inference intègre désormais le support Gaudi d'Intel directement dans TGI, avec une architecture multi-backend et compatibilité Gaudi1/2/3. L'article présente les bénéfices (diversité matérielle, coût, production-ready features) et liste les modèles optimisés (Llama 3.x, Mistral, Mixtral) et donne un démarrage rapide (Docker + exemple curl) pour lancer TGI sur Gaudi. Il mentionne aussi le FP8 via Intel Neural Compressor et invite à tester et contribuer.
DeepSeek publie Open R1: Update #4, version améliorée de DeepSeek-V3 avec licence MIT et de meilleures capacités d’instruction et de code. Les benchmarks montrent des gains (MMLU-Pro 81.2, GPQA 68.4, AIME 59.4, LiveCodeBench 49.2) et des progrès en développement web, écriture chinoise et appel de fonctions. Utilisation via Inference Providers (Fireworks, Hyperbolic, Novita) et Text Generation Inference, avec des spéculations sur le finetuning (pré-entraînement et post-formation).