Local / Open Source
Hugging Face
Updates
- Feature1d ago· GitHub Releases
Release v5.24.0
Hugging Face Transformers 5.24.0 adds support for Inkling 2 Mini, a multimodal mixture-of-experts model with configurable loop passes. It also introduces Apertus 1.5, which accepts text, images, and audio and generates text.
- Feature9d ago· GitHub Releases
Release v5.23.0
Hugging Face Transformers 5.23.0 adds support for Inkling 2 Small, a multimodal mixture-of-experts model accepting text and images. The model has 594 billion total parameters, 17 billion active parameters, and supports multiple natural and programming languages.
- Feature16d ago· GitHub Releases
Patch release v5.22.2
Hugging Face Transformers 5.22.2 hardens Jinja sandboxing and fixes distributed execution by setting device_id only when using NCCL. The patch targets security and distributed-computing behavior.
- News23d ago· GitHub Releases
Patch release v5.22.1
Hugging Face Transformers 5.22.1 removes unfinished Ml4 documentation. The patch contains no reported code changes.
- Feature24d ago· GitHub Releases
v5.22.0: Mistral Large 4 (Le Chonk)
Hugging Face v5.22.0 adds Mistral Large 4, a multimodal mixture-of-experts model with roughly 1 trillion total parameters, 52 billion active parameters, and a 1 million-token context window. It supports reasoning, coding, agent workflows, and more than 160 languages.
- Feature24d ago· GitHub Releases
Patch release v5.21.1
Hugging Face v5.21.1 fixes the rope_scaling configuration setter. The patch restores remote-code loading in vLLM after it was broken in the prior release.
- Feature27d ago· GitHub Releases
Release: v5.21.0
Hugging Face Transformers 5.21.0 adds support for MiniCPM-OCR, a document-understanding vision-language model, and Flint, a mixture-of-experts language model using hybrid attention and learned attention sinks.
- FeatureOct 19· GitHub Releases
Patch release v5.20.2
Hugging Face Transformers v5.20.2 fixes input type loss during generation that could break multimodal models. It also addresses zero expert gradients with bfloat16 expert parallelism, incorrect VLM modality declarations, and an unnecessary host-to-device copy.
- FeatureOct 16· GitHub Releases
Release v5.20.1
Hugging Face Transformers v5.20.1 adds support for Reflection AI’s Beam model, designed for coding and agentic workloads on user-managed infrastructure. The patch also fixes model loading on older PyTorch versions and issues affecting batched generation, expert parallelism, and distributed checkpointing.
- FeatureOct 15· GitHub Releases
Release v5.20.0
Hugging Face Transformers 5.20.0 adds support for new models including Meta’s Muse Spark Open and InclusionAI’s Ling 3.0. Muse Spark Open is a multimodal 816B-parameter MoE model with 42B active parameters and a context window exceeding one million tokens.
- FeatureOct 6· GitHub Releases
Release v5.19.0
Hugging Face Transformers 5.19.0 adds EmbeddingGemma 2 support for multimodal embeddings across text, images, audio, and video. The release also standardizes router-logit output behavior for mixture-of-experts models, introducing a breaking change.
- FeatureSep 30· GitHub Releases
Release 5.18.0
Hugging Face Transformers 5.18.0 adds support for Nemotron 3 Diarization, an open-weight model for streaming and offline speaker identification. It handles up to eight speakers and offers configurable latency and frame resolution.
- FeatureSep 9· GitHub Releases
Release 5.17.0
Hugging Face 5.17.0 adds HYV4 support, including the 780B-parameter Hy4-Preview mixture-of-experts model with 49B active parameters and a 1-million-token context window. The architecture uses latent and sparse attention, attention sinks, and independent hyper-connections.
- FeatureAug 26· GitHub Releases
Release v5.16.1
Hugging Face v5.16.1 adds GLM-5.3-Flash support alongside several small fixes. GLM-5.3-Flash is a natively multimodal mixture-of-experts model with 320 billion total parameters, 18 billion active parameters, and hybrid sparse and linear attention.
- FeatureAug 26· GitHub Releases
Release: v5.16.0
Hugging Face v5.16.0 adds support for the experimental Qwen4-Exp model architecture. The architecture combines GatedResidual, Qwen Sparse Attention, Per-Layer Embedding, and linear attention components for more efficient long-context inference.
- FeatureAug 19· GitHub Releases
Patch release: v5.15.1
Hugging Face v5.15.1 fixes DFlash and MTP candidate generation issues, accelerator image processing with Lanczos filtering, and Gemma 4 video device handling.
- NewsAug 11· Changelog
Hugging Face changelog updated
Hugging Face added an IBM Research post titled “Thinking of ACE? We Can Do It with Fewer Tokens.” The excerpt provides no further technical details about the method or results.
- FeatureAug 10· GitHub Releases
Release: v5.15.0
Hugging Face Transformers 5.15.0 adds support for Meta’s Apache-2.0-licensed Muse Glimmer, a 30B multimodal model intended for local agentic applications. It also adds GraniteSWA and GraniteMoeSWA support.