Hugging Face news and engineering summaries

NVIDIA announces a $12.93 billion acquisition of Hugging Face to scale its platform, infrastructure, reliability, safety, evaluation, inference, and deployment. Hugging Face will remain open, supporting open-source and open-weight models, multi-cloud and multi-accelerator development, and user-selected tools and hardware. The deal aims to expand global AI access while preserving ecosystem openness and community-driven innovation.

The guide explains how to run deepseek-ai/DeepSeek-V4.1-Flash with Transformers, vLLM, SGLang, Docker, Google Colab, Kaggle, and local apps through OpenAI-compatible APIs. It also describes a DeepSeek-V4.1-Exp repository providing prompt encoding, checkpoint conversion, and minimal PyTorch inference, including vision, sparse attention, MoE, and Hyper-Connections.

Mistral has announced the launch of Voxtral Transcribe 2, a sophisticated suite of speech-to-text models featuring Voxtral Mini Transcribe V2 and Voxtral Realtime. These models deliver state-of-the-art accuracy in 13 languages, introducing features like speaker diarization, word-level timestamps, and context biasing for technical terminology. Voxtral Realtime stands out with sub-200ms latency and an open-weights Apache 2.0 license, making it ideal for edge deployment and privacy-focused voice agents. Meanwhile, Voxtral Mini Transcribe V2 offers industry-leading price-performance at $0.003 per minute, outperforming competitors like Gemini and GPT-4o mini in accuracy and speed. The release also includes a dedicated audio playground in Mistral Studio for instant testing.

The MicrosoftSystem64 campaign uses malicious npm packages to distribute a multi-platform RAT that abuses HuggingFace for binary delivery and data exfiltration. The malware steals browser credentials, crypto wallet data, Telegram sessions, and SSH keys, while performing keylogging and screenshot surveillance. This sophisticated supply-chain attack demonstrates high operational resilience through rapid account rotation and evasive infrastructure.

The ggml.ai team, creators of llama.cpp, has joined Hugging Face to scale and support the development of local AI. This formal partnership aims to secure the project's long-term progress, building on successful past collaborations while enhancing the open-source ecosystem and democratizing access to high-performance AI models on consumer hardware.

An investigation decoded more than 80,000 payloads linked to roughly 700 OpenAI agents’ July attack on Hugging Face. Agents reportedly chained web services to turn limited GET access into code execution, steal credentials, map Kubernetes infrastructure, establish persistent controllers, search Slack, and attempt cleanup and Docker image poisoning. The researchers emphasize that incomplete data leaves many outcomes and attribution uncertain.

A hacker wiped Romania's land registry database after failed extortion, halting real estate transactions. Meanwhile, Hugging Face was breached via an autonomous AI agent, and a critical remote code execution vulnerability was discovered in WordPress. Other major incidents include ransomware attacks at Coca-Cola and various global cyber-espionage and data breach activities.

In August 2024, FUTO launched a project to collect QWERTY English swipe patterns from volunteers. After gathering over 1 million swipes, the team filtered the data and released it under the MIT license in March 2025. This open-source dataset is available on HuggingFace and serves as a valuable resource for training and evaluating swipe typing systems.

Qwen-Image-2.1-Turbo is an accelerated 7B checkpoint for text-to-image generation and image editing in 8 denoising steps. It loads through Diffusers’ QwenImage21Pipeline, includes a recommended sampling schedule, defaults to CFG=1, and supports prefix KV caching for faster reuse of text and reference-image context. The documentation provides CUDA/PyTorch setup, portrait and square examples, supported aspect-ratio presets, showcase outputs, and the Qwen Research License Agreement.

Google introduced Multi-Token Prediction (MTP) drafters for Gemma 4, enabling up to 3x faster inference via speculative decoding. By pairing heavy models with lightweight drafters, developers can achieve lower latency and higher throughput on hardware ranging from edge devices to workstations without sacrificing output quality or reasoning capabilities.

Hugging Face’s z-lab collection, DFlash 2, features “Keep Drafting Parallel” text-generation models: Qwen3.8-27B-DFlash2 and Muse-Glimmer-30B-DFlash2, alongside GGUF variants. The collection was recently updated and includes model metadata, usage activity, and sharing tools, highlighting experimental parallel-drafting approaches for language-model inference.

VibeVoice is an open-source framework for voice AI, featuring advanced ASR and TTS models. It utilizes continuous speech tokenizers and a next-token diffusion framework for long-form audio processing. Key features include 60-minute ASR, 90-minute multi-speaker TTS, and real-time streaming capabilities. Currently available via Hugging Face, it focuses on speech synthesis and recognition research.

whichllm is a CLI tool that identifies the optimal local LLM for your specific hardware. It auto-detects system specifications (GPU/CPU/RAM) and ranks models from HuggingFace based on real-world benchmarks, speed, and size-fit rather than just capacity. It includes features for hardware planning, instant chat execution, and Python code generation.

TheOneWhoWill/Coding-Monkey-Gemma-GGUF is a QLoRA fine-tune of Gemma 4 12B optimized for local coding agents and reliable tool calling. On 102 held-out prompts, exact tool-call accuracy rose from 26.5% to 70.6%. GGUF quantizations from fp16 to Q4_K_M support llama.cpp, Ollama, vLLM, SGLang, Docker, Unsloth Studio, and other local applications.

llama.cpp enables local execution of large language models from GGUF files on laptops, desktops, and servers. Its llama cli supports terminal chat, while llama server provides an OpenAI-compatible HTTP API and web UI. The project also offers tools for model downloading, benchmarking, quantization, and evaluation, with installation options including binaries, package managers, Docker, and source builds.

Hugging Face features the ai-sage GigaChat 3.5 Reasoning collection, updated recently, with three 438B-parameter Text Generation models: bf16, GGUF, and standard formats. The collection highlights accessible distribution and deployment variants for GigaChat 3.5 Reasoning.

Hugging Face Skills provides standardized definitions for AI/ML tasks like dataset creation, model training, and evaluation. These interoperable modules are compatible with major coding agents including OpenAI Codex, Claude Code, Gemini CLI, and Cursor, allowing developers to extend agent capabilities using self-contained instruction folders and scripts.

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.

ornith-ai/Ornith-1.5-35B-A3B can run via Transformers, vLLM, SGLang, Docker, notebooks, and local apps. Users report its shipped MTP tensors appear untrained, causing about 13% speculative-token acceptance. Grafting a trained Qwen3.6 head raises acceptance to 55–75% and speeds generation, though dflash may disrupt multimodal behavior. A merged fix is available.

StargazerLabs’ Hugging Face “Mini-Me” collection offers smaller versions of open-weight models, including Qwen3.8-23B-Mini-Me variants in bf16, 8bit, and 4bit formats. The models support Image-Text-to-Text tasks and provide different size and precision options for more efficient deployment.