DiffusionGemma uses parallel text diffusion instead of sequential token generation, achieving 1000+ tokens/sec on H100 GPUs with trade-offs in output quality.
Google DeepMind releases Gemma 4 12B, a 12-billion-parameter model with unified vision and audio processing that runs on 16GB consumer hardware.
Hugging Face hackathon project demonstrates how tiny language models can power real-time simulations that frontier models cannot economically support.
JetBrains' new Mixture-of-Experts model achieves 2x speedup over dense peers while activating just 2.5B parameters per token.
Liquid AI unveils a sparse 8-billion-parameter model with 1-billion active parameters, trained on 38T tokens—a scale comparable to frontier model training runs.
NVIDIA releases diffusion language models at 3B, 8B, and 14B scales that generate and refine tokens in parallel, offering latency improvements for GPU-constrained inference workloads.
Stability AI releases four new audio models capable of generating full-length songs, with open-weights tiers and licensing deals backing the release.
Allen Institute releases OlmoEarth v1.1, a more efficient earth-observation model family that maintains v1 performance while reducing compute through shorter token sequences.