Coding Tools
Transformers adds native GGUF support for Apple Silicon
Hugging Face Bloghigh confidence
Why it matters
- →Developers can now run quantized GGUF models directly in Python without external conversion steps.
- →Apple Silicon users get near-native llama.cpp performance within the standard Transformers API.
- →The integration simplifies prototyping and evaluation of local models using familiar PyTorch hooks.
Hugging Face has integrated native GGUF support into its Transformers library, enabling direct loading of quantized models via the standard from_pretrained API. The implementation leverages llama.cpp’s ggml kernels to deliver inference performance on Apple Silicon that closely rivals the dedicated llama.cpp runtime. Initial support targets Qwen3.5 architectures, allowing developers to experiment with quantized checkpoints using familiar PyTorch tools and evaluation workflows. This move consolidates local AI development by combining the accessibility of GGUF formats with the extensibility of the Hugging Face ecosystem.
Read originalMore from Hugging Face Blog
© Hugging Face BlogModels & Labsother
Jun Kim joins Hugging Face to lead MLX ecosystem
Hugging Face is professionalizing the MLX ecosystem by hiring Jun Kim, creator of oMLX, to work full-time on Apple Silicon AI. This move signals a serious commitment to stabilizing local inference outside NVIDIA's CUDA monopoly. By integrating oMLX into Hugging Face’s infrastructure, they aim to streamline the conversion of Hugging Face models into MLX-compatible formats. It removes the bottleneck of volunteer maintenance for a framework that is becoming essential for Mac-based AI development.
Hugging Face Blog
© Hugging Face BlogResearchresearch
AISI releases verified evals for GPT-5 and Claude Opus 4
The UK AI Security Institute has published verified benchmark results for GPT-5 and Claude Opus 4 using EvalEval’s standardized schema, solving the reproducibility crisis in frontier model testing. By releasing raw configuration data alongside scores from benchmarks like SWE-Bench Pro and Humanity's Last Exam, they prove that inference-time compute drastically alters performance curves. This moves evaluation beyond opaque leaderboards into auditable science, allowing researchers to see exactly how protocol choices skew reported capabilities. It sets a new standard for transparency in high-stakes AI security assessments.
Hugging Face Blog
© Hugging Face BlogResearchmodels
Physics-based pruning beats heuristics for LLM compression
Multiverse AI reframes block removal as an Ising glass optimization problem, capturing the hidden couplings between transformer layers that mean-field methods ignore. By mapping block importance to spin interactions via a Hessian matrix, they turn model compression into a search for low-energy states rather than independent block scoring. This approach yields a massive 23-point MMLU gain over existing baselines when compressing Llama-3.3-70B by half, proving that many-body physics tools can unlock deep compression without retraining. The method scales to large models using classical and quantum-inspired solvers, offering a rigorous alternative to heuristic pruning.
Hugging Face Blog
More in Coding Tools
Coding Toolscoding
Claude Code v2.1.280 adds Opus 5.5 and fixes agent stability
Claude Code just got its first major model upgrade with Opus 5.5 as the new default, bringing a 1M context window and aggressive pricing that reshapes local inference economics. Beyond the headline model swap, this release quietly stabilizes the background subagent system, fixing critical issues where tool lists were rebuilt instead of cached and reports were silently lost during compaction. The UI layer also sees significant polish, with mouse support in fullscreen mode and fixes for Windows terminal rendering that had plagued power users. This is less about new features and more about making the agent runtime reliable enough for heavy, multi-step workflows.
Claude Code Releases
Coding Toolscoding
llama.cpp b11104 adds multi-addr server binding
The llama-server now binds to multiple addresses, a practical upgrade for anyone running local inference behind reverse proxies or complex network setups. This change removes the previous single-address limitation, allowing flexible routing without external workarounds. While the release includes standard binaries for CUDA 13 and ROCm 10.0, the networking feature is the real differentiator here. It makes self-hosted deployments slightly more robust for power users who need granular control over traffic flow.
llama.cpp Releases
Coding Toolscoding
llama.cpp b11105 adds ROCm 10 and Snapdragon support
This release quietly cements llama.cpp as the universal inference runtime by finally bringing full ROCm 10.0 support to both Linux and Windows, closing a long-standing gap for AMD GPU users who previously had to rely on workarounds or older versions. The inclusion of CUDA 13 builds alongside CUDA 12 ensures compatibility with the latest NVIDIA driver stacks without forcing users into beta territory. Perhaps most notably, the addition of native Snapdragon support on Linux marks a significant step toward efficient AI inference on ARM-based mobile and edge devices, expanding the hardware ecosystem beyond traditional x86 and NVIDIA dominance.
llama.cpp Releases