Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 5, 2026, 9:00 AM
Hugging Face has launched the Inference Endpoints managed service, supporting one-click deployment of open-source LLMs such as Falcon, LLaMA, and StarCoder; the article uses Falcon 40B-instruct as an example.
The official Hugging Face blog published a tutorial introducing how to use Node.js combined with open-source models such as WizardCoder-15B to achieve streaming generation from text to web apps via the Inference API or self-hosted Inference Endpoints.
A Hugging Face blog post analyzes the inconsistency between MMLU scores on the Open LLM Leaderboard and those in the LLaMA paper, empirically comparing them with the original UC Berkeley implementation.
A Hugging Face blog post introduces how to fine-tune the Massive Multilingual Speech (MMS) model using adapters to achieve automatic speech recognition (ASR) for low-resource languages.
The official Hugging Face blog announced support for deploying Livebook notebooks to Hugging Face Spaces to build a voice-only chat application based on the Whisper model; this application can be set up and deployed within 15 minutes, demonstrating Livebook's capabilities in multi-model serving and cluster distribution within the Elixir ecosystem.
The Hugging Face blog announced the release of models optimized via Core ML 6-bit palettization for four official versions: Stable Diffusion 1.4/1.5/2-base/2.1-base.
OpenAI announced multiple API updates, including more controllable models, function calling capabilities, longer context windows, and lower prices.
Hugging Face published a blog post systematically comparing the performance of human annotators and GPT-4 in preference evaluation, using outputs from open-source models such as Vicuna-13B, Koala-13B, Oasst-12B, and Dolly-12B across 327 instruction prompts.
A new feature on Hugging Face Hub allows users to execute SQL queries on any public dataset stored on the platform using DuckDB; all public datasets have been automatically converted to Parquet format, with file URLs accessible via the /parquet endpoint, enabling direct queries of remote Parquet files when combined with DuckDB's httpfs extension.
The Technology Innovation Institute (TII) of Abu Dhabi released the Falcon series of open-source large language models, including Falcon-40B and Falcon-7B, both under the Apache 2.0 license.
The official Hugging Face blog published a complete tutorial on integrating speech recognition into Unity games, using its Unity API to convert microphone input to text in real time; the tutorial includes runnable code for scene setup, UI component configuration, 10-second WAV recording encoding implementation, API calls, and UI status feedback.
Hugging Face has released a dedicated LLM inference container (DLC) for Amazon SageMaker, built on Text Generation Inference (TGI).
A Hugging Face blog post introduces how to optimize Stable Diffusion for CPU using the OpenVINO Neural Network Compression Framework (NNCF) and 🤗 Optimum.
Hugging Face and Microsoft have partnered to launch a natively integrated Hugging Face Hub model catalog in Azure Machine Learning Studio, supporting one-click real-time inference deployment for thousands of Transformers models; this feature is now available in public preview across all regions that support Azure Machine Learning.
Hugging Face's official blog released the QLoRA paper, proposing an efficient method for fine-tuning a 65B parameter model on a single 48GB GPU via 4-bit NormalFloat quantization.
Hugging Face, in collaboration with EleutherAI and Stability AI, commissioned Trail of Bits to complete an external security audit of the safetensors library. The audit confirmed no critical vulnerabilities leading to arbitrary code execution. This Rust-implemented tensor serialization format will gradually become the default save and load format for the transformers library; it is currently a core dependency, with default enablement planned within months.
A Hugging Face blog post proposes applying instruction-tuning methods to Stable Diffusion to enhance its ability to follow specific image editing instructions (such as cartoonization, deraining, denoising, and deblurring).
OpenAI has released the ChatGPT iOS app, which supports conversation synchronization and voice input, and integrates the latest model improvements.
Hugging Face and Intel jointly released Q8-Chat, a ChatGPT-style chat demo running on a single Intel Sapphire Rapids CPU (32 cores, batch size=1).
The Hugging Face blog announced the integration of the RWKV architecture into the transformers library. This architecture merges the memory efficiency of RNNs with the long-range dependency modeling capabilities of Transformers.
Hugging Face released a tutorial guiding users to run the 4-bit quantized Vicuna-13B model on AMD GPUs that support ROCm (such as RX6900XT, MI210).
Hugging Face has proposed a new decoding method called assisted generation, which uses smaller assistant models (such as pythia-160m) to pre-generate candidate token sequences.
Hugging Face released a tutorial demonstrating how to train the RoBERTa-base model from scratch using TensorFlow and TPU, including training a Unigram tokenizer on WikiText v1.
Hugging Face has released a new feature supporting the direct conversion of Apache Spark DataFrames into Hugging Face Datasets via the Dataset.from_spark() method, avoiding intermediate Parquet file writes and reads. Processing time for 16GB of data dropped from 22 minutes to 12 minutes, representing a speedup of over 40%. This feature has been merged into the Hugging Face Datasets codebase and was contributed by Databricks.