Hugging Face AI Updates: August 26, 2026

1. IBM Releases Granite 4.2 Dense Reasoning Language Models

IBM. IBM published Granite 4.2 on the Hugging Face blog, its first family of dense reasoning language models with explicit chain-of-thought and switchable thinking, non-thinking, and low-effort modes. The release includes 3B, 8B, and 30B parameter models under Apache 2.0, pre-trained on roughly 15 trillion tokens with context extended to 512K. The 8B and 30B variants received additional agentic training to call tools, run code, and use terminals in sandboxed environments. Source

2. Quantization-Aware Healing Produces a 4-Bit Model That Beats Its Full-Precision Original

Multiverse Computing. A Multiverse Computing post on the Hugging Face blog introduced Quantization-Aware Healing, a technique that distills a compressed, quantized model from the original full-size teacher instead of a degraded checkpoint. Applied to a GPT-OSS 120B model compressed to 60B parameters and quantized to 4-bit MXFP4, the result outperformed its own bfloat16 version on 7 of 9 benchmarks, including gains of 7.4 points on long-context reasoning and 5.6 points on AIME 2025 math. The method reached peak performance in about 100 training steps versus roughly 700 for competing quantization-aware training. Source