IBM has released Granite 4.2, a family of dense, decoder-only reasoning large language models in 3B, 8B, and 30B sizes. These models are pre-trained from scratch on approximately 15 trillion tokens and feature a context window extended to 512K tokens.
Multiverse Computing introduces Quantization-Aware Healing (QAH), a method that distills compressed, quantized large language models directly from their original pre-compression teachers. Applied to a GPT-OSS 120B model compressed to 60B parameters and quantized to MXFP4, the approach produces a model that beats its own full-precision bfloat16 version on 7 of 9 benchmarks.
The UK AI Security Institute (AISI) has made publicly reported evaluation methods and findings available through EvalEval's Evaluation Cards platform to improve benchmark reproducibility. This release includes verified results, context, and configuration information for five specific benchmarks: HealthBench, FrontierMath, Humanity's Last Exam, SWE-Bench Pro, and Terminal-Bench 2.0.
Hugging Face's Transformers library now supports loading GGUF quantized models, leveraging the underlying ggml kernels to bring performance close to llama.cpp. This integration allows developers to run quantized checkpoints directly within the standard PyTorch-based API on supported hardware.
Jun Kim, the creator and maintainer of oMLX, has joined Hugging Face to support the MLX community. This move aims to provide stability and faster development for the open-source project by transitioning it from a side job to a fully maintained and funded initiative.
The tokenizers library has released a version 1 release candidate that significantly improves performance for text encoding, addressing bottlenecks in training and serving workflows. The update maintains API compatibility with v0.23 while delivering substantial speedups across ten model families.
Multiverse has released a paper detailing a method that reframes large language model pruning as a constrained binary optimization problem mapped to an Ising glass. By treating transformer blocks as spins with pairwise couplings derived from the loss Hessian, the approach identifies optimal block removal configurations without requiring iterative benchmarking.
Researchers have introduced "consistency guidelines" within the ALTK-Evolve system, a new mechanism designed to address the variability in LLM agent performance that standard average accuracy metrics often hide. By identifying and stabilizing decision points prone to flipping, this approach significantly improves task success consistency without sacrificing overall capability.
TRL v1.14 introduces LoRA support to the AsyncGRPOTrainer, allowing it to train a Low-Rank Adaptation adapter and synchronize only that small file to vLLM inference workers. This architecture decouples training and generation jobs on separate machines by using a shared Storage Bucket as a filesystem bridge, eliminating the need for NCCL or direct network communication between nodes.
Gradio has released Workflow1111, a project that rebuilds most of AUTOMATIC1111's stable-diffusion-webui feature set using the gr.Workflow framework. The application consists of a single canvas containing eleven media pipelines built from seventy-three nodes, covering text-to-image, image-to-video, and various preprocessing tasks.
IBM has released Granite Time Series PatchTST-FM-r2, a ~385M-parameter time series foundation model that achieves state-of-the-art zero-shot performance on the GIFT-Eval leaderboard under permissive commercial licenses.
A Multiverse Computing paper introduces a method to train language models to refuse only specific harmful subsets of a topic, rather than the entire category, addressing the limitations of blunt topic-level guardrails. The approach uses boundary-aware self-distillation with coverage repair and in-distribution benign data to maintain safety while reducing false refusals on legitimate prompts.
Researchers introduce NeoMME, a family of 260M and 800M multilingual multimodal encoders that process text and raw image patches using a single bidirectional Transformer. Unlike generative visual language models, NeoMME does not rely on separate pretrained vision towers or causal decoders, training the entire model from scratch with a masked discrete-diffusion objective.
Hugging Face has released funes, a tool that transforms existing coding agent session logs into a durable, searchable memory layer for Claude Code, Codex, pi, and Hermes. It indexes traces locally using vector and BM25 search, allowing agents to retrieve exact provenance without external services.
A guide demonstrates fine-tuning the Liquid AI LFM2.5-350M model using Group Relative Policy Optimization (GRPO) to enhance structured output generation. The process involves training on a subset of the NVIDIA Nemotron dataset and evaluating performance via the IFStruct benchmark.
An engineer has open-sourced a reproduction of Surya Narreddi’s viral project, which trains a coding model to paint watercolours using JavaScript and reinforcement learning. The implementation utilizes the TRL library and OpenEnv to create an end-to-end pipeline on Hugging Face for training, scoring, and inference.
IBM and Confluent have integrated IBM's Granite Time Series Foundation Models (TSFM) into Confluent Cloud, enabling real-time forecasting and anomaly detection directly within Apache Flink.
Researchers introduce BenchMIRT, a method for auditing large language model benchmarks at the individual prompt level using multidimensional Item Response Theory. By analyzing performance across 100 models and 34,000 questions, the tool disentangles mixed capabilities like safety and general reasoning that often obscure benchmark scores.
Hugging Face has released @huggingface/kernels, a library for loading and running optimized WebGPU kernels from the Hugging Face Hub, alongside an initial collection of 207 kernels. The release also introduces Fleet, an in-browser GPU benchmarking suite that crowdsources performance and correctness evidence from users' devices.
The Open ASR Leaderboard has introduced its first Global South languages with two new evaluation sets: Monsoon en-IN (Indian English) and Monsoon hi-IN (Hindi). These additions address the lack of demographic diversity in current benchmarks, which have historically focused on European languages.