Hugging Face
1 Tool · 170 Intel
Hugging Face is an American machine learning company, founded in 2016 in New York City by Clément Delangue, Julien Chaumond, and Thomas Wolf. It began as a teen-focused chatbot app before pivoting to become a widely used open-source platform and community for sharing machine learning models, datasets, and tools, including its popular Transformers library. It is often described as functioning like a "GitHub of machine learning" for researchers, data scientists, and engineers.
Is this you? Sign in with X to claim this profile.
Tools
Open-source ML-engineer agent from Hugging Face that reads research papers, trains models, and ships them end to end.
Intel
NVIDIA and AI4Bharat describe Nemotron-pretraining-Indic-v1, about 2.8 trillion synthetic tokens across 22 Indian languages, and the open pipeline for translation, quality scoring, and repair. The dataset is still listed as releasing soon.
Photoroom opens up the data pipeline behind its PRX image model: why pre-training favors coverage over prettiness, assembling and profiling mixed sources in Lance, re-captioning everything with a VLM, writing Mosaic streaming shards with aspect-ratio bucketing, then filtering and deduplicating.
Tencent open-sourced EVIE-4.5B, a late-interaction visual document retrieval model with elastic Prefix-MRL embeddings (64-2048D from one projection) and HAC, a training-free token-compression scheme cutting index size roughly 20x while leading ViDoRe V3.
Nex AGI ships Nex-N2.5-mini, a multimodal agentic model refining computer-use, browsing, and visually grounded task execution from the Nex-N2 line at a smaller footprint than the flagship Max.
Dharma-AI previews VocabX, which specializes a small model's vocabulary around per-token cost instead of frequency. On a Portuguese essay grader built on Qwen3-4B, they report a 77% inference cost cut without changing the base model. Full paper still pending.
Nvidia's Nemotron 3.5 Content Safety evaluates prompt, image and response as one context, enforces a custom policy you supply rather than a built-in taxonomy, and can emit reasoning traces for audit. Twelve trained languages plus zero-shot transfer across ~140.
Hugging Face engineers show how TRL's AsyncGRPOTrainer now trains and syncs only LoRA adapters, not full models, across separate HF Jobs machines via a storage bucket and routing proxy, cutting a 500-step run from 3h27m to 53min.
Every Eval Ever and Hugging Face Community Evals now interoperate: one JSON schema records who ran an evaluation, on which model, via what access path, with which generation settings and metric definition, and a converter posts those records onto model pages. Roughly 229k results, 22k models, 31 sou
Liquid AI releases an experimental speculative-decoding draft model for LFM2.5-VL-3B, detailing the shared-representation architecture and showing up to 3.13x faster on-device decoding and 2.66x on H100 without quality loss.
Interfaze releases interfaze-1.1-lite, an open-weight model for deterministic tasks such as OCR, diarized transcription and classification. A reasoning core dispatches specialist models and emits schema-conforming JSON with boxes, timestamps and confidences, on one 80 GB GPU.
Hugging Face discloses an intrusion driven end to end by an autonomous agent system: a malicious dataset abused two code-execution paths in dataset processing, then the actor took node-level access, harvested credentials and moved laterally. Public models and packages verified clean; rotate your tok
Liquid AI releases DSpark draft models for its LFM2.5 family, including a 24B MoE drafter. On a B300 with an FP8 target in SGLang, throughput averages 2.4x at concurrency 1 and 1.7x at 256, with per-benchmark tables.
Tencent released WeVisDoc-2B, an Apache-2.0 document parser fine-tuned from Qwen3-VL that converts page images into structured Markdown with LaTeX and HTML tables, topping OmniDocBench v1.6 and PureDocBench among compared end-to-end parsers.
Hugging Face's PEFT team measures how completely LoRA has taken over parameter-efficient fine-tuning — around 98% of PEFT model cards and 95% of image-generation checkpoints — then walks through when the alternatives actually win and how to decide.
Hugging Face launches a Space that reports, per Hub model, whether it runs unmodified on AWS Trainium through the torch-neuronx backend. It lists core counts, wall time and memory, and provides run scripts plus a public compile cache.
This Hugging Face blog post explains how the SVDQuant/Nunchaku 4-bit quantization method (weights and activations, W4A4) is now natively supported in the Diffusers library, letting users load pre-quantized diffusion transformer checkpoints via from_pretrained() without a separate CUDA inference engine. It reports concrete gains such as ~1.7s image generation at ~12GB peak memory versus ~24GB for BF16, and introduces a companion 'diffuse-compressor' toolkit for quantizing and publishing new architectures.
NVIDIA releases Kumo Relational, a few-shot foundation model that predicts churn, demand or fraud from database tables with no training or feature engineering. It reports beating supervised pipelines on 41 tasks, available as an open library and a NIM.
ServiceNow shows deep-research agents leak private context through their outbound web queries via the mosaic effect, and that optimizing purely for task performance increases it. Their privacy-aware RL method lifts strict chain success 48.7%→58.7% while cutting leakage 34%→9.9%.
funes indexes the session traces your coding agents already produce into a local Lance dataset and exposes recall and get tools back to the agent. Retrieval fuses vector and BM25 ranking, reranks with a cross-encoder, weights by recency, and returns original text with provenance.
Hugging Face rebuilt its Kernels project: a dedicated 'kernel' repo type on the Hub, loading restricted to trusted publishers by default with an explicit opt-in flag otherwise, code signing, reproducible Nix builds, and revamped CLIs.
IBM releases Granite Switch 4.1 (8B and 30B), composing a decision adapter, a guardian adapter and a fast ASR model into one model using activated LoRAs that reuse KV cache.
A migration guide that keeps GitHub Actions as the CI front end but runs jobs on Hugging Face infrastructure via ephemeral self-hosted runners, cutting Trackio's CPU CI time about 30% and enabling a test suite on real CUDA hardware.
H Company releases Holo4, a 27B/35B-A3B open computer-use agent that operates GUIs, code, MCP, and APIs through one interface, scoring 61.7% on OSWorld 2.0 versus 81.8% for Opus 5.5 at far lower cost, with full trajectories released for inspection.
Sentence Transformers v6.0 adds MultiVectorEncoder for ColBERT-style late-interaction retrieval and the training stack around it. The walkthrough finetunes a domain multi-vector model in 14.5 hours on a single RTX 3090 that outperforms every dense, sparse and lexical retriever tested on medical data