AI BriefingHuggingFace Blog news
The latest AI and developer news from HuggingFace Blog, with short summaries.
Visit website →NEWAi2 releases XYBench, a benchmark for testing LLMs' ability to identify and correct user misconceptions
Developed by Ai2 and partners, XYBench contains 8,115 questions to measure how well models recognize when a user's request is based on a misconception and steer them toward the underlying goal.
NEWllama.cpp and ggml reach v1.0.0 with stable API and ABI guarantees
The release establishes strict semantic versioning rules for the C API and expands distribution channels to include Docker, Debian, and conda-forge.
NEWIBM Releases LLMON Markup Language and Granite 4.1 Preview Models with Enhanced Prompt Injection Robustness
The experimental preview models use reserved special tokens to separate instructions from data, achieving higher robustness against prompt injections without sacrificing general benchmark performance.
NEWHugging Face Releases Fine-Tuning Guide for TRELLIS.2 3D Generation Models
A new script and container allow users to fine-tune TRELLIS.2 with LoRA on custom 3D assets, significantly improving geometric accuracy for specific domains like footwear.
Tether AI Research releases BotPsy, a 0.8B parameter agentic model outperforming larger rivals
The open-source model achieves SOTA tool-use performance for its size, scoring 45.0 on average compared to 38.8 for Qwen3.5-2B-Think.
Token Pooling-Aware Training for Multi-Vector Embedding Models
The method uses Ward hierarchical clustering and an averaged loss function to compress document vectors, achieving a 97.8% retention rate on NanoBEIR at a 4x compression factor.
Hugging Face Hub Now Available on Google Cloud Marketplace for Unified Billing
Organizations can now link their Hugging Face account to Google Cloud to consolidate billing and avoid new vendor onboarding.
Hugging Face Introduces Decision Models for Probabilistic, Non-Generative AI Outputs
The new interface supports choice, score, and noul primitives with a unified API across transformers and llama.cpp.
NVIDIA and Elith Release Nemotron-Personas-Finance-Japan, a 9M-Record Synthetic Dataset for Japanese Financial AI
The CC BY 4.0 licensed dataset contains 50 fields per record and is built using NeMo SDG-PGMs and gemma-4-31B-it to support privacy-compliant financial AI training.
Amazon releases HOPE for higher-order pruning of MoE language models
The method uses second-order interactions to maintain performance during aggressive expert removal, outperforming first-order baselines on Qwen3.5 and GLM-4.5 models.
AllenAI Releases ToyBlocks Simulation Framework for Robot Learning and Evaluation
The open-source framework pairs automatic data generation with benchmarks showing VLA models generalize poorly while frontier agents solve 30/30 stacking tasks.
Open ASR Leaderboard adds Egyptian and Nigerian English private data via Liva AI
New private splits from Liva AI include 1.34 hours of Egyptian and 1.55 hours of Nigerian conversational speech to improve evaluation reliability.
NVIDIA Releases Cosmos 3 Open Models for Interactive World Simulation
The 16B parameter checkpoints support multi-sensor driving generation and bimanual robot action prediction under the OpenMDW1.1 license.
Gradium Details Phonon TTS: On-Device Speech Model Tops CPU Streaming Benchmarks
Gradium's Phonon TTS, available in 31m and 40m sizes, tops the Hugging Face Open TTS Leaderboard's CPU streaming benchmark for time-to-first audio.
Dharma Field Extractor Outperforms Frontier LLMs on Document Field Extraction
The 2B parameter model achieves a 0.834 field-level F1 score, surpassing Claude Opus 5 by 1.8 points and GPT-5.6 Luna by 9.3 points.
StateBench: AI2 releases benchmark showing frontier models fail on long-context state tracking
AI2 introduces StateBench, a four-level benchmark revealing that while recurrent architectures generalize better than transformers, top models like Gemini 3.8 Flash and Claude Opus 5.5 still fail on hard state-tracking tasks.
IBM Research Palette: 20B open-weight model ranks second on PresentBench, first in Economics and Education
A fine-tuned 20B model using a designer-coder split and check-repair harness achieves a 59.7 mean score on PresentBench, ranking second overall and first in Economics and Education.
IBM Research CUGA Outperforms Hermes and Deep Agents on AppWorld Benchmark
IBM's open-source CUGA agent harness achieved an 81.1% task completion rate on the AppWorld benchmark, significantly outperforming Hermes and Deep Agents using the same GLM-5.3-flash model.
Collinear AI releases CWE-bench v2 with 100 public security tasks
The benchmark covers 91 CWEs with deterministic scoring, where the top model still fails 19% of tasks.
Hugging Face Introduces Model Integration Contracts to Catch LLM Regressions Early
The new standard uses small CPU fixtures to verify model behavior on PRs, catching a UMT5 attention bug that persisted for two months.