Hugging Face AI Updates: October 16, 2026

1. Hugging Face Ships Day-Zero Support for Meta’s Muse Spark Open

Hugging Face. Hugging Face published integration support for Muse Spark Open, Meta’s open-weight multimodal mixture-of-experts model with about 816B total parameters, about 42B active per token, and a context window of up to 1M tokens. The model takes text, image, audio, and video input, produces text, and uses 256 experts per MoE layer with 8 selected per token. Checkpoints are available in BF16 and NVFP4 (Muse-Spark-Open-816B-A42B and Muse-Spark-Open-816B-A42B-NVFP4), and the post covers inference with Transformers, llama.cpp, SGLang, vLLM, and MLX, deployment on Inference Endpoints, and fine-tuning with TRL. Source

2. FLUX.3 Image Klein Lands in Diffusers With Bounding-Box Layout Control

Hugging Face. Black Forest Labs’ FLUX.3 Image Klein is supported in Diffusers from launch, with two open checkpoints: a timestep-distilled model (16 steps, no CFG) and a base model (50 steps, CFG). One model handles text-to-image, multi-reference image-to-image, and a new layout control mode where bounding boxes in the prompt specify where regions appear or which regions of a reference image move. Hugging Face reports the distilled checkpoint generates a 1024 x 1024 image in about 2 seconds on an H100, NVFP4 and FP8 weights load via from_single_file, and LoRA training scripts are provided for text-to-image and editing. The pipeline ships as a Modular Diffusers pipeline, so blocks can be removed or added. Source

3. IBM Releases Granite Switch 4.1 Models Combining Chat, Decisions, and Speech on the Hub

Hugging Face. IBM published Granite-Switch-4.1-8B and -30B on the Hub, each composing a base model with a general Decision aLoRA, a Guardian aLoRA for safety decisions, and the Granite-Speech-5.0-Turbo-CTC ASR model behind a single endpoint. Because the adapters are Activated LoRAs, the KV cache from a base model generation can be reused for later adapter-based decisions, and IBM says the composed model uses roughly the same VRAM as the base model alone. Users can also compose custom Granite Switch models from adapters in Granite Libraries, and a Speech-to-Decisions Space demonstrates the 8B model transcribing, summarizing, and screening audio in real time. Source

4. Hub Tensor Viewer Loads Large Sharded Models Up to 20x Faster

Hugging Face. When a user opens a model’s model.safetensors.index.json, the Hub now sends the headers of all shards in a single request instead of the browser fetching each shard individually. Hugging Face reports DeepSeek-V3.2 now loads in 0.45s instead of 9.2s, GLM-5.3 in 0.57s instead of 4.0s, and Qwen2.5-72B-Instruct in 0.27s instead of 1.2s. Source