Introducing Falcon ASR
The Technology Innovation Institute released Falcon-ASR, a 1.6 billion parameter speech recognition model focused on Arabic and the Emirati dialect, reporting 20.92% average word error rate across six Arabic test sets.
First-party releases and research from Hugging Face Blog. Headlines and excerpts link to the original articles.
The Technology Innovation Institute released Falcon-ASR, a 1.6 billion parameter speech recognition model focused on Arabic and the Emirati dialect, reporting 20.92% average word error rate across six Arabic test sets.
NVIDIA teams fine-tuned Nemotron 3 models with supervised fine-tuning, reinforcement learning, and feedback-driven inference to reach gold-medal-level results at IOI 2026 and IMO 2026. The IMO result was graded by official IMO graders; the IOI result was an unofficial, unsupervised benchmark not included in the official ranking.
The Technology Innovation Institute released Falcon-Emirati-7B, an Emirati-dialect Arabic model built on Falcon-H1-Arabic. It scores 84.83% on the Alyah benchmark and 85.57% on the UAE portion of ArabCulture-Dialogue, leading models many times its size in comparisons the institute ran.
Microsoft's ThinkingBox, now available through Hugging Face, grades AI agents on the database state and side effects they leave behind rather than their responses, running each of 507 business workflows 20 times. The joint blog reports that many cleanly terminated, seemingly successful agent runs still failed executable state checks.
Ai2 has open-sourced AstaBrief 8B, a report-generation model built on Qwen3-8B that turns a research question and retrieved literature into a cited report. It is used as Fast mode in the Asta platform and released with its training data and example workflow.
ServiceNow CoreAI introduced AutoSynthData, a pipeline that uses a target model's failures and a stronger teacher's successes to generate validated training tasks for enterprise agents. In EnterpriseOps Gym experiments, synthetic supervised fine-tuning improved Gemma-4-26B-A4B-it mean Pass@1 by 7.2 percentage points in Hybrid and from 18.77% to 27.18% in ITSM.
Hugging Face released Olmo-core 3, an open training framework for large mixture-of-experts (MoE) models, designed to scale MoE training into the trillion-parameter range while preserving computational efficiency. It is part of the systems behind the next generation of Olmo.
Hugging Face has introduced the Open TTS Leaderboard, an objective-metric evaluation for open-source and multilingual text-to-speech and voice cloning models. It measures intelligibility, speed, and speaker similarity, and includes a Listen tab for human comparison and voting.
Hugging Face Hub has added a dedicated filter and tagging system for reinforcement learning environments, hosted as dataset repositories. Four environment frameworks are registered as dataset libraries, and tagging requires no new repo type, registry, or sign-up.
Hugging Face published details on tokenizers v1, a release candidate focused on performance. The post says v1 encodes text 3 to 30 times faster than v0.23 with one thread on an Apple M4 Max while producing the same token IDs.
A Hugging Face blog post introduces the Consistency Analyzer and consistency guidelines in ALTK-Evolve, a method to measure and reduce agent unreliability. On AppWorld, a ReAct agent using GPT-4.1 succeeded on 77.4% of runs on average but on all five runs for only 53.0% of tasks, a 24.4-point gap that consistency guidelines cut to 12.0 points.
Hugging Face describes an async GRPO setup using TRL v1.14's AsyncGRPOTrainer with LoRA, where a trainer Job and two vLLM Jobs share a Storage Bucket instead of NCCL. A proxy adds auth headers, routes rollouts by KV prefix, and broadcasts adapter loads; five runs cut 500-step training from 3h27m to 53m.
A Hugging Face blog post describes Workflow1111, a Gradio Workflow rebuild of much of AUTOMATIC1111's stable-diffusion-webui feature set as one canvas of 73 nodes and eleven media pipelines. Each output node becomes a REST endpoint and, with mcp_server=True, an MCP tool.
Hugging Face announced NeoMME, a family of 260M and 800M multilingual multimodal encoders that process text tokens and raw image patches in a single bidirectional Transformer trained from scratch. Checkpoints are released under the Apache 2.0 license, with a fine-tuned NeoMME-Retriever for visual document retrieval.
Hugging Face has published funes, an open-source local memory layer for coding agents such as Claude Code, Codex, pi, and Hermes. It indexes existing session traces, supports recall and ask commands, and can optionally sync memory to a private Hugging Face dataset you own.
A Hugging Face blog post describes an open reproduction of a project that trains a coding model to paint watercolours by writing p5.brush JavaScript, using TRL and OpenEnv with three reward mixes. The blog reports trained models and datasets published openly on the Hub.
A Hugging Face guide describes fine-tuning LiquidAI's LFM2.5-350M with GRPO via TRL, improving IFStruct structured-output scores from 22.6% to 29.7% in about 100 steps and roughly 500 samples. The run fits a free-tier Colab or Kaggle GPU, and evaluation ran locally through llama.cpp on a MacBook.
Researchers introduced BenchMIRT, a multidimensional Item Response Theory method for auditing LLM benchmarks at the level of individual prompts. Trained on results from 100 LLMs across 16 benchmarks and more than 34K questions, it independently recovered safety and general reasoning as dominant dimensions.
Hugging Face released @huggingface/kernels, a JavaScript library for loading and running optimized WebGPU kernels from the Hugging Face Hub, along with 207 Apache-2.0 licensed kernels and Fleet, a browser-based GPU benchmarking tool.
Voice Arena and Hugging Face added Hindi and Indian English evaluation sets, called Monsoon, to the Open ASR Leaderboard, marking its first Global South language. The sets include public and private splits across 4,888 speakers and introduce OIWER for Hindi orthographic variation.