The best of dev and AI, scored every day by an agent. Filtered, summarized, ranked. No color, no noise — just the substance.
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.
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).
This blogpost shows how to finetune a reranker (Cross Encoder) with Sentence Transformers for domain-specific retrieval, detailing datasets, loss functions, training arguments, evaluators, and the trainer. It demonstrates that small, domain-tuned models can outperform public rerankers and scale with larger bases, in a two-stage retrieve-and-rerank setup.
Gradio released a major update to gr.Dataframe, adding multi-cell selection, row numbers, pinned columns, and new UI controls like a copy button and fullscreen view. The update also improves accessibility, styling, and supports static columns, search, and filtering for more intuitive data exploration. Gradio notes 70 issues resolved in six weeks and provides upgrade instructions and a quick example to get started.
The article introduces updates to the Inference Endpoints analytics dashboard, highlighting real-time metrics, faster data loading, and customizable time ranges with auto-refresh. It also adds a replica lifecycle view to monitor state transitions across multiple replicas, aiding monitoring and debugging. Feedback is invited as the features evolve.
The post explains how to run OlympicCoder-7B locally using LM Studio and a VS Code extension. It walks through installing LM Studio, loading OlympicCoder-7B (4-bit GGUF), and wiring LM Studio to VS Code with Continue.dev, enabling code completion, generation, explanation, and tests. OlympicCoder is open-source and optimized on the CodeForces-CoTs dataset, not a Claude-like assistant.
Hugging Face submitted its response to the White House AI Action Plan, arguing that open models and open science are foundational to AI success. They propose three recommendations: recognize open source as fundamental, prioritize efficiency and reliability to enable broad adoption, and secure AI through open, traceable, and transparent systems. They illustrate with examples like OlympicCoder beating Claude 3.7 with 7B parameters and open OLMo 2 models.
At GTC 2025, NVIDIA unveiled three open-source releases for physical AI: Cosmos Transfer (a 7B World Foundation Model with multicontrols for controllable world scenes), the Open Physical AI Dataset, and Isaac GR00T N1, the first open humanoid-reasoning foundation model. Cosmos Transfer uses per-sensor ControlNets and adaptive spatiotemporal maps to generate photorealistic sequences with controllable layout and motion.
Xet migration moved roughly 6% of Hugging Face Hub download traffic to Xet storage, validating the architecture under real-world usage. Unlike LFS that deduplicates at the file level, Xet uses content-defined chunking (~64 KB) to deduplicate at the byte level, so edits affect only changed chunks and large files transfer faster. The system combines a Xet-aware client, a content-addressed store with an LFS bridge, and S3 for final storage, all geared toward production readiness.
Gemma 3 is Google's open, multimodal LLM family released in 1B, 4B, 12B, and 27B variants with up to 128k context and support for 140+ languages. The 4B/12B/27B models are multimodal (images+text) using SigLIP, with pan-and-scan inference and memory-optimized KV cache; the 1B remains text-only. It can also be used as a text-only LLM by omitting the vision encoder.
Open R1: Update #3 reports progress on reproducing the code-reasoning aspects of DeepSeek-R1, including a new CodeForces-CoTs dataset with ~100k samples in C++ and Python. It introduces the IOI benchmark and OlympicCoder models (7B and 32B), showing that training on CodeForces-CoTs yields top-tier performance, with OlympicCoder-32B outperforming several open-weight models. The post also covers dataset construction, benchmarks, and practical tips for scaling and prompting, such as prefilling with