Loading…
Hugging Face
AI platform and community providing machine learning models, datasets, libraries, and tools for building AI applications.
- 319 posts summarised
- latest
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
From their blog
Latest articles
HiRO-ACEv2: Simulating extreme weather at 3 km resolution for a fraction of the cost
HiRO-ACEv2 generates global 3 km climate data for extreme weather at roughly 1/300th the cost of the X-SHiELD model, with even larger savings for single-region runs. The system pairs a coarse 100 km AI model (ACE) with the new HiROv2 downscaler that simultaneously predicts rain, wind, and sea-level pressure instead of just rainfall, and has improved skill at reproducing extreme events. When fed coarsened X-SHiELD data HiROv2 matches the target’s high-resolution statistics within 3% for events as rare as one in a million data points, and accurately estimates tropical cyclone energy for storms never seen during training. Over a 10-year simulation the full system’s hurricane energy and extreme event snapshots compare well with X-SHiELD. However, HiRO cannot yet handle other climates because it depends on X-SHiELD-specific ACE tuning; future work includes training on warmer-world datasets and adding temperature and snow outputs.
Run fast, private agents on-device with LFM2.5-24B-A2B
LFM2.5-24B-A2B is a 24-billion-parameter reasoning model with only 2 billion active parameters per token, designed for fast, private on-device agents. It uses a Mixture-of-Experts architecture with 40 layers, 64 experts per block, and top-4 routing, combined with gated short convolutions and GQA. The model produces an explicit chain of thought before answering and supports tool calling and multi-step tasks. Pre-training scaled to 33 trillion tokens, followed by mid-training and a five-stage post-training pipeline including supervised fine-tuning, multi-domain distillation, and agentic reinforcement learning. On benchmarks, it leads or matches larger models on tool use and instruction following, achieving 48.09 on LiveClaw and 96.75 on IFEval, while decoding 184 tokens per second on an Apple M5 Max.
Impactful scheduling for GPU clusters
Ai2 replaced its priority-based GPU cluster scheduler with a system combining GPU time budgets, hierarchical fair-share allocation, and a time-slicing contract to prioritize high-impact research while maintaining full occupancy. The previous priority scheduler induced GPU squatting, priority inflation, and constant negotiation over non-preemptable workloads, exemplifying a tragedy of the commons. In the new model, managers allocate proportional GPU time budgets to projects, making resource allocation a transparent administrative process rather than an operational debate. The scheduler tracks occupancy over a time window with weights set by budgets, and time-slicing enforces minimum runtime protection before workloads become preemptible. However, interactive sessions suffered from the 8-hour cap, prompting investment in a CPU-only cluster and restorable sessions, while potential capacity fragmentation for large workloads remains under investigation.
The model that didn't exist, so you made it yourself
The author wanted a small version of the prompt rewriter that ships with Qwen-Image 2.1, but the official 9B model requires about 20 GB of memory. Using ML Intern in HuggingChat, they created a 0.8B version that runs on a CPU, returns valid output 99.7% of the time, and uses about a quarter of the teacher’s tokens. The compute for the entire project, including having the 9B model label 8,797 example requests, came to USD 16. Over a few days, the author made five more models using the same approach, each starting as a message in HuggingChat with ML Intern switched on. The author’s prompts include verified facts, a baseline request, a smoke test, and a spending cap. The six models built include a citrus disease vision model, a Huggy character LoRA, a camera-angle LoRA, a doodle-in LoRA, a pocket rewriter, and a distilled text-to-image model.
Introducing Falcon ASR
The Technology Innovation Institute in Abu Dhabi released Falcon-ASR, a 1.6 billion parameter speech recognition model focused on Arabic, especially the Emirati dialect, while also supporting English, French, Spanish, and Portuguese. On the Open Universal Arabic ASR Leaderboard, Falcon-ASR achieved a 20.92% average word error rate across six test sets, outperforming the best published average of 23.17% in the reference snapshot. The model recorded a 22.73% word error rate and 10.19% character error rate on an internal Emirati evaluation, 4.07% better in word error rate than the next best system compared. Training data included background noise, overlapping speech, music, reverberation, telephony effects, speed, and pitch variations to handle diverse recording conditions. Falcon-ASR builds on the Falcon3-Audio architecture and provides word-level timestamps linking transcribed words to audio positions.
One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO
Starting from Nemotron 3, the team applied supervised fine-tuning (SFT) and reinforcement learning (RL) to create specialists for competitive programming and olympiad mathematics. At the IOI, the Nemotron-3-Ultra-CC model, trained with SFT on 22,000 curated problems and paired with GenCorrect iterative refinement, scored 535.4 of 600 points, exceeding both the gold threshold and the top human score. At the IMO, a combination of SFT and RL checkpoints scored 30 of 42 points by generating, critiquing, and refining natural-language proofs without external tools. Both projects followed a shared recipe: start with a strong base, curate domain-specific data, apply standard post-training, and tie the model to an inference loop that improves candidates. The results show that specialization and test-time compute together produce gold-level performance, and all models, data, and inference pipelines are available on Hugging Face.
Falcon-Emirati: When an LLM Learns the Dialect, the Culture, and the Nuance
Falcon-Emirati-7B is a dialect-specialized language model built on Falcon-H1-Arabic to understand and generate Emirati Arabic, a Gulf dialect with distinct vocabulary and cultural context that Modern Standard Arabic models miss. The model uses a hybrid architecture combining Mamba state-space models and Transformer attention running in parallel inside every block, with fused outputs for efficient long-context processing. Training required a dedicated data pipeline of authentic Emirati web content, MSA material about Emirati culture, and synthetic data guided by strict glossaries and style rules to cover gaps in available dialectal text. On the Alyah benchmark, Falcon-Emirati-7B scored 84.83%, outperforming larger multilingual models, and achieved 85.57% on the UAE cultural understanding task in ArabCulture-Dialogue. The 7B parameter size was chosen as the best balance of quality against training and inference cost for dialect adaptation.
The Agent Said It Was Done. The Database Disagreed.
ThinkingBox, a benchmark from Microsoft and Hugging Face, grades AI agents on the database state and side effects they leave behind rather than on tool calls or final responses. In a common-set ablation covering 121,680 valid trials across 12 LLM models, 79,853 attempts failed executable checks, and 67.24% of those failures still terminated cleanly and invoked a state-changing tool without error, with wrong field values in 77.61% of them. The benchmark runs every task 20 independent times from identical clean backends, reporting pass@1, pass@20, and observed 20/20 consistency to separate capability from reliability. For example, Kimi-K3 solved 93.89% of tasks at least once but only 13.41% consistently, while Claude Opus 5 completed 47.53% of tasks on every attempt, and Claude Opus 5.5 gained no additional dependability over its predecessor despite higher pass@1. ThinkingBox is available through Hugging Face’s OpenEnv with MIT-licensed code and CDLA-Permissive-2.0 benchmark data.
Open-sourcing AstaBrief, the fast report-generation model in Asta
AstaBrief is an open-weights 8B-parameter model developed by Ai2 for generating cited scientific reports from a research question and literature excerpts, designed for evidence-grounded outputs. It was trained from Qwen3-8B using supervised fine-tuning on 47K filtered real user queries and direct preference optimization, focusing on data quality. Unlike Asta’s Thinking mode, which uses a multi-step pipeline, AstaBrief writes the full report in a single pass to reduce generation time. In Asta’s Fast mode, it averages 51.1 seconds per report versus 178.5 seconds for Thinking mode, roughly 3.5× faster, while competitive on answer and citation quality. The model and training data are open-sourced on Hugging Face, enabling institutions to deploy behind their own firewall and supporting future open scientific models.
AutoSynthData: Generating Training Data for Enterprise Agents
AutoSynthData, built by ServiceNow CoreAI, generates training data for enterprise agents by targeting capability gaps from evaluation, using a stronger teacher such as Qwen3.8-27B to demonstrate successes. It uses a target model’s failures and a stronger teacher’s successes to decide what to learn next. The system produces tasks composing a system specification, user prompt, and verifier, ensuring feasibility, realism, and difficulty against the target environment. After sample-level checks and batch-level review, the dataset grows via target and multiply phases, generating 2,000 samples in 18 hours for the Hybrid domain. In EnterpriseOps Gym, synthetic SFT on Gemma-4-26B-A4B-it improved mean Pass@1 by 7.2 points (35% relative) in Hybrid and from 18.77% to 27.18% in ITSM, showing the approach works across domains.
Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs
Olmo-core 3 is an open training infrastructure upgrade for large mixture-of-experts (MoE) models, designed to scale to trillion parameters while preserving efficiency. It switches from FSDP to a DDP-based system that keeps experts resident on GPUs and routes data to them, avoiding weight gathering overhead. The system combines expert, pipeline, and distributed optimizer parallelism with optimizations like rowwise expert parallelism, GPU-resident routing, and grouped GEMM. In a benchmark, increasing experts from 8 to 128 while routing each token to four experts kept active parameters at ~3.2B, total parameters grew from 4.6B to 47B with throughput falling less than 5%. The infrastructure supports MXFP8 lower-precision format, achieving about 21% higher throughput than BF16, and has been tested on a 1.2-trillion-parameter model.
Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning
The open-source TTS model landscape has exploded, with over 8,000 models on Hugging Face Hub, yet evaluation methods remain unstandardized and cannot scale. The Open TTS Leaderboard addresses this by using objective metrics such as word error rate (WER), speaker similarity (SIM), and speed (RTFx and TTFA) to evaluate models in hours instead of weeks. It supports multilingual evaluation and voice cloning, providing separate tabs for listening to outputs and comparing streaming performance. Unlike arena-based leaderboards that significantly underrepresent open models due to high hosting costs, this leaderboard prioritizes open-source models. The project encourages community participation through feedback and voting; evaluation scripts will be open-sourced soon, and future updates will incorporate new datasets, languages, and metrics.
NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction
NVIDIA Kumo Tabular is an open foundation model for tabular classification and regression that predicts labels in a single forward pass without training, tuning, or feature engineering. It uses a Transformer architecture with column, row, and in-context attention, including length-aware attention temperature to maintain sharp attention on large tables. Pretrained entirely on artificial tables generated from structural causal models with realistic imperfections such as missing values, coarsened features, and heavy-tailed targets, it comes in three sizes (28M to 215M parameters) and handles missing values without imputation. On the TabArena, BeyondArena, TALENT, and ScoringBench benchmarks, Kumo Tabular achieves top rankings and establishes a new accuracy-efficiency frontier, running 17x faster than LimiX-2. The model is available on Hugging Face under the OpenMDW-1.1 license and runs via NVIDIA’s open-source structured-data-models library.
Getting the Source Right, Not Just the Fact: Source-Aware Verification for MCP Agents
MCP agents can produce claims that are factually true but attributed to the wrong source, a failure known as cross-source conflation that source-blind checkers miss. ProvenanceGuard preserves source identity by decomposing answers into claims, routing each to its relevant tool output, and checking support and attribution separately, using local models like MiniLM and DeBERTa. On medical agent traces with 281 interactions, it caught 138 of 139 claims experts flagged, scoring 0.802 F1 on blocking decisions and correct source 86% of the time. It flagged 67 supported claims for review under a cautious policy and added roughly half a second per answer in local tests. Blocked answers can be fixed through a repair loop for rewriting or fallback, avoiding unverifiable responses while keeping sensitive traces in a controlled environment.
Holo4: powering generalist computer-use agents
Holo4 is a new series of agentic models designed to interact with software through any available interface: GUIs, code, MCP, and APIs. It comes in two sizes, a 27B dense model and a 35B-A3B Mixture of Experts model, both available on the H Models API. The models were trained through supervised and reinforcement learning on environments and tasks generated by the Agentic Task Factory, which produces verifiable tasks from documentation alone. On the OSWorld 2.0 benchmark, Holo4 27B scores 61.7% against 81.8% for Opus 5.5, achieving competitive performance at a fraction of the cost. The company also released Holotron4 Nano, an updated version of Holotron 3 built on the Nemotron 3 Nano Omni model using the same post-training stack.
Welcome RL Environments to the hub
Hugging Face Hub now has a dedicated RL Environments filter for dataset repositories tagged with rl-environment, addressing the problem of siloed environments across frameworks like Harbor, Verifiers, OpenEnv, and NeMo Gym. Instead of building a new registry, the Hub uses existing dataset repos with framework tags that generate loading snippets, so an environment consists of tasksets stored as datasets and runtimes provided by frameworks. The environment gives an agent a task, responds to actions with observations, and scores the outcome with rewards for evaluation or training, and users can run environments by pulling the dataset and using the framework’s runtime and verifier code. The release includes pre-tagged environments such as BeyondSWE, Terminal-Lego, Multi-SWE-RL-Verified, and others, with run commands provided for each framework. Future plans include per-config snippets and automatic framework tagging on upload, and the Hub already stores data, versions it, gates it, and serves it to millions.
Accelerating vision-language models with LFM2.5-VL-DSpark
Liquid AI released an experimental DSpark draft model for their vision-language model LFM2.5-VL-3B. The drafter adds a speculative decoding path that trades a minimal increase in memory footprint for a larger speedup without changing output quality. The 280M-parameter drafter, a 4-layer attention-only transformer, captures the target model’s hidden states at fixed tapped layers to draft a block of candidate tokens. On an M5 Max, decoding runs 2.30x to 3.13x faster by task, with end-to-end latency improvements of 1.56x to 2.62x. The drafter ships with day-one support for llama.cpp, MLX-VLM, and SGLang, and is available on Hugging Face in Safetensors and GGUF formats.
How to Use NVIDIA Warp and MjWarp to Accelerate Robotics Simulation and Learning Workflows
NVIDIA Warp is a Python framework for writing GPU-accelerated kernels that compiles statically typed Python to CUDA. MuJoCo Warp (MJWarp) implements MuJoCo physics on Warp, enabling batched simulation of hundreds to thousands of independent robot environments on NVIDIA GPUs. The core API transition is small: mjw.put_model and mjw.put_data replace mujoco.MjModel and mujoco.MjData, and mjw.step advances the entire batch. Proper throughput measurement requires GPU synchronization and warm-up to avoid timing queue latency instead of actual GPU work. The post demonstrates moving an SO-101 follower arm from a single CPU world to 2,048 parallel MJWarp environments, validating parity and measuring aggregate world-steps per second.
**Know Who Spoke When: Build Real-Time, Multi-Speaker AI with NVIDIA Nemotron 3 Diarization**
NVIDIA Nemotron 3 Diarization is an open-weight, 100M-parameter model that ranks #1 on VoiceArena’s Diarization-Bench leaderboard with a 14.72% Diarization Error Rate (DER). It classifies who spoke when in conversations, handling overlapping speech and supporting up to eight speakers across live and recorded audio. The model uses a Sortformer approach with arrival-order speaker labeling, a 31-layer Transformer encoder, and two memory mechanisms for streaming inference. It accepts 16 kHz single-channel audio and outputs a floating-point tensor of speaker-activity probabilities at a configurable time step. Developers can combine its timestamps with automatic speech recognition to produce speaker-attributed transcripts, with recommended input-buffer latencies ranging from 0.32 to 30.4 seconds.
How UK AISI and EvalEval Are Making Benchmark Results Reproducible
The UK AI Security Institute (AISI) has begun using EvalEval’s Every Eval Ever (EEE) schema and Evaluation Cards platform to openly share LLM evaluation results. This collaboration, stemming from a joint workshop at NeurIPS 2025, aims to make benchmark results more reproducible by providing standardized configuration and context alongside verified scores. AISI released results for five benchmarks—HealthBench, FrontierMath, Humanity’s Last Exam, SWE-Bench Pro, and Terminal-Bench 2.0—across six frontier models, plus two cyber evaluations using a partially overlapping model set. The release accompanies a paper studying how inference-time compute and evaluation protocol affect performance, revealing that models continued solving tasks when given correctness feedback. By adopting a shared reporting schema, evaluators can compare findings across studies and understand how setup choices influence reported performance, while EvalEval invites model developers, evaluation creators, and researchers to contribute to the shared infrastructure.
Topics they write about
All topics →Weekly digest
The best engineering reads, once a week.