The best of dev and AI, scored every day by an agent. Filtered, summarized, ranked. No color, no noise — just the substance.
Ettin introduces the first SoTA paired encoder-only and decoder-only models (17M–1B params) trained identically for apples-to-apples comparisons. It extends the ModernBERT recipe to both architectures, with encoders beating ModernBERT and decoders beating Llama 3.2 and SmolLM2, while preserving architecture-specific advantages.
Hugging Face has deployed Xet as the Hub’s storage backend, migrating hundreds of petabytes and millions of repos with minimal disruption. The migration relies on a Git LFS Bridge and background content migrations to support both LFS and Xet, allowing a no-hard-cutover transition. The approach aims to scale storage with AI workloads while preserving existing workflows.
Kimina-Prover-72B uses a Test-Time Reinforcement Learning (TTRL) search to autonomously discover and reuse lemmas for long-horizon Lean proofs, plus an error-fixing module that interprets Lean messages for targeted corrections. It achieves a state-of-the-art miniF2F performance (92.2% pass rate) and shows pass@1/32/1024 of 63.9/84.0/87.7, with two distilled variants released (8B and 1.7B).
Asynchronous inference decouples action prediction from execution in robotic policies, reducing runtime lag and enabling replanning with action chunks. The article describes a two-component architecture (PolicyServer and RobotClient) using gRPC to achieve ~2× speedups and continuous operation, and explains why sequential inference falls short.
ScreenEnv is a Python library that runs isolated Ubuntu desktop environments in Docker to test and deploy GUI agents with full desktop control, including window management and file operations. It supports both the Model Context Protocol (MCP) and a Direct Sandbox API, enabling flexible integration with existing backends or AI systems. The article provides setup examples and a quick path to build a custom Desktop Agent using smolagents.
Building the Hugging Face MCP Server enables customized AI Assistants to access the Hub and thousands of apps through a single URL. The article compares MCP transports (Streamable HTTP vs SSE) and outlines three patterns: Direct Response, Request Scoped Streams, and Server Push Streams, with their trade-offs and needed connection management. It also covers making the server dynamic and remotely configurable, plus client-side usage hints (TypeScript/Python).
Hugging Face and AMD optimize custom kernels for MI300X to boost Llama 3.1 405B inference in FP8 on 8 GPUs. They combine fused residual/RMS norm/FP8 conversion, SwiGLU, and a Skinny GEMM kernel, with benchmarks and open-source tooling in hf-rocm-kernels.
Reachy Mini is an open-source, programmable robot for human-robot interaction and AI prototyping. It runs Python (and soon JS/Scratch), ships as a kit, and starts at $299, with a wireless compute version at $449. It emphasizes open hardware/software, a growing community, and privacy controls by design.
The article introduces the Model Context Protocol (MCP) and explains how MCP servers can extend LLMs with new abilities via an app-store-like ecosystem. It guides readers to find MCP-compatible spaces in Hugging Face Spaces and demonstrates wiring a Flux.1 Kontext[dev] MCP server to an LLM to edit images from text prompts.
Hugging Face outlines three major alerts that bolster production infrastructure monitoring. The highlighted NAT Gateway throughput alert acts as an early warning for unusual egress traffic and helps with cost optimization, while other alerts cover Kubernetes API request errors, rate limiting, and a bonus alert for new clusters sending zero metrics.
SmolLM3 is a competitive open 3B model (11T tokens) with 128k long context via NoPE and YaRN, and multilingual support across six languages. It uses a Llama-based decoder with Grouped Query Attention, intra-document masking, and training-stability tweaks, plus a dual Instruct/Reasoning model.
Efficient MultiModal Data Pipeline identifies data pipeline waste in multimodal training, like idle GPUs and excessive padding. It outlines a five-stage approach—from visualizing the dataset to packing batches with a knapsack strategy—to reduce padding and maximize throughput. The post also introduces an iterable dataset and a repository for practical reproduction.
Annonce de la compétition NeurIPS 2025 E2LM visant à capturer les signaux en phase précoce d'entraînement des LLMs dans le domaine scientifique. Les submissions seront évaluées via ScoreSQ, ScoreRC et ScoreCS, avec Score = α1·ScoreSQ + α2·ScoreRC + α3·ScoreCS et des poids α1=0.5, α2=0.1, α3=0.4. L'inscription se fait sur Hugging Face et l'évaluation s'appuie sur lm-evaluation-harness, avec des phases allant de juillet à novembre 2025 et résultats prévus le 04 novembre 2025.
Sentence Transformers can finetune sparse encoder/embedding models for retrieval, hybrid search, and reranking. The post outlines training components (model, datasets, losses, trainer, evaluators) and provides practical examples, including using naver/splade-v3 and decoding sparse embeddings for interpretability.
Le NVIDIA Llama Nemotron Nano VL est un modèle Vision-Language 8B optimisé pour l’identification et l’OCR de documents, offrant une haute précision OCRBench v2 et une extraction fiable de texte, tableaux et éléments visuels. Disponible sur Hugging Face, il s’appuie sur Llama-3.1-8B-Instruct et C-RADIOv2-VLM-H et peut être post-entraîné via NVIDIA NeMo; un tutoriel Notebook permet de construire rapidement des solutions d’automatisation pour factures, contrats et documents de santé.
Gemma 3n is now available in major open-source libraries (transformers, timm, llama.cpp) and runs on-device with multimodal support (text, image, audio, video). It ships two sizes (5B/8B real; 2B/4B VRAM), a flexible MatFormer architecture, and mix-and-match sub-models. It includes a MobileNet-V5 vision encoder and a USM audio encoder, supports 140 languages, and offers a Hugging Face Space demo.
SGLang now supports Hugging Face transformers as a backend, enabling high-throughput, low-latency inference for any transformers-compatible model. It can auto-fall back to transformers or be forced with impl='transformers', and adds features like RadixAttention for efficiency. The article includes practical code snippets and deployment patterns (offline engine, server, and OpenAI-compatible API).
New post shows efficient fine-tuning of FLUX.1-dev on consumer hardware via QLoRA with the diffusers library, targeting peak VRAM under ~10 GB on a single GPU. It explains loading a quantized 4-bit base model, training FP16/BF16 LoRA adapters, uses an 8-bit AdamW optimizer, and discusses options to load or merge LoRA adapters with results demonstrated on an RTX 4090.
Groq is now a supported Inference Provider on the Hugging Face Hub, enabling serverless inference on model pages and tight integration with Python and JavaScript SDKs. The LPUs promise lower latency and higher throughput for LLMs, with support for models like Meta's Llama 4 and Qwen's QWQ-32B, and two billing modes: direct provider key or routing through HF.
The article analyzes how long prefill prompts can block the prefill queue in a multi-request setting, and explains that decoding steps are light but must be sequential. It discusses two patterns - chunked prefill and request-parallel prefills - and why long prompts undermine throughput, with implications for vLLM scheduling.