Hugging Face AI Updates: September 23, 2026

1. Transformers Loads GGUF Checkpoints Directly and Lands Close to llama.cpp Speed

Hugging Face. GGUF quantized checkpoints now load through from_pretrained(model_id, gguf_file=filename) and run through the normal generate() path, with transformers serve exposing them behind an OpenAI-compatible endpoint. Supported variants include Q6_K, Q5_K_M, and Q4_K_M, which for a 4B Qwen model come in at 3.53 GB, 3.14 GB, and 2.74 GB against a 8.42 GB BF16 reference. On a MacBook Pro M2 Max, Transformers measured close to llama.cpp across all three checkpoints after generation loop optimizations cut synchronization overhead and ggml Metal kernels were integrated for quantization, normalization, and attention. Source

2. UK AISI Is Publishing Its Evaluation Results Through EvalEval’s Open Schema

Hugging Face. The UK AI Security Institute is using EvalEval’s infrastructure to publish evaluation results in a reproducible form, releasing verified results, context, and configuration for the five benchmarks in its main paper through Evaluation Cards. The problem being addressed is concrete: results are reported across incompatible formats without enough information to reproduce them, and rerunning an evaluation can cost more than the original. AISI feedback also shaped the Every Eval Ever schema, which sits alongside the Institute’s OptStop and HiBayES work on evaluation efficiency and statistical rigor. Source

3. oMLX Maintainer Jun Kim Joined Hugging Face, Project Stays Apache 2.0

Hugging Face. Jun Kim, creator and maintainer of oMLX, joined Hugging Face to work on the MLX community. The project remains Apache 2.0 with Kim still leading it, moving from side project to funded and maintained. The stated focus area is streamlining the path from a Transformers model definition to a reference MLX implementation that different engines can consume, with existing collaboration across mlx-lm, mlx-vlm, and LMStudio. Source