Hugging Face Blog, summarized in your digest

Ai2 replaced its priority-based GPU scheduler with GPU time budgets, hierarchical fair-share allocation, and a time-slicing contract to prioritize impactful research while keeping clusters occupied. The system gives projects guaranteed capacity shares, charges protected workloads against budgets, and preempts/requeues jobs after declared minimum runtimes; unallocated time remains interruptible. Across a 30-day test, teams received 98% of owed GPU hours and occupancy stayed at 98%; debug-job p90 wait fell from two hours to 30 seconds, while human-required repairs dropped 74%.

ML-intern enabled the creation and publication of several custom models from natural-language prompts, handling planning, dataset generation, training, evaluation, and deployment on Hugging Face hardware. Projects included a citrus-disease VLM improving accuracy from 14.9% to 52.8%, image LoRAs, a CPU-ready 0.8B Pocket Rewriter, and a 4-step Agate model. Budgets were enforced with baselines and smoke tests; total compute across the projects was about USD 103.

Liquid AI released two open decision models: d1-3B and experimental d1-omni-600M. Built on Liquid Foundation Models, they answer in one forward pass rather than generating tokens. d1-3B accepts text and images, leads models under 10B on Decision Index 0.2.1 with 48.57, and achieves a mean benchmark score of 82.9. d1-omni-600M supports text plus image or audio, scores 78.4 with 600M parameters, and d1-3B runs in 16–50 ms on Jetson edge devices. Both are open-weight on Hugging Face.

Technology Innovation Institute introduced Falcon-ASR, a 1.6B-parameter speech recognition model focused on Arabic, especially Emirati, while also supporting English, French, Spanish, and Portuguese. It achieved 20.92% average WER across six Arabic test sets, 22.73% WER and 10.19% CER on an internal Emirati evaluation, and 5.74% WER across seven English test sets. The model handles varied recording conditions, provides word-level timestamps, and uses one set of weights without language flags.

Nemotron 3 was fine-tuned into specialist systems that achieved gold-level results at IOI 2026 and IMO 2026. The IOI system, Nemotron-3-Ultra-CC with SFT and GenCorrect, scored 535.4/600, above the 361.12 gold threshold and top human score. The IMO system combined general, SFT, and RL checkpoints in a generate-verify-refine loop, scoring 30/42, above the official 29-point threshold. The results show that curated reasoning data, post-training, and feedback-driven inference work together; models, datasets, benchmarks, and pipelines are available on Hugging Face.