Ideogram-4-FP8 and Gemma-4-12B Models Now Available on Hugging Face

Three new models launched on Hugging Face within hours of each other: Ideogram AI's FP8-quantized image generation model and Google's Gemma-4-12B language models.

Ideogram-4-FP8

Ideogram AI released ideogram-ai/ideogram-4-fp8, an FP8-quantized version of its Ideogram-4 image generation model. FP8 quantization reduces model size and inference memory requirements while maintaining image quality, making the model more accessible for deployment on consumer hardware.

The base Ideogram-4 model is known for text rendering capabilities in generated images. Pricing and specific benchmark details for the FP8 variant have not yet been disclosed.

Google Gemma-4-12B Models

Google released two versions of Gemma-4-12B:

• google/gemma-4-12B: Base pre-trained model
• google/gemma-4-12B-it: Instruction-tuned variant optimized for chat and task completion

The 12-billion parameter models represent a mid-size option in Google's Gemma family. Specific context window size, benchmark scores, and training cutoff dates were not provided in the initial Hugging Face model cards.

Both variants are released under Google's Gemma license, which allows commercial use with certain restrictions.

Technical Details

All three models are hosted on Hugging Face's model hub, making them accessible through the Transformers library and compatible inference frameworks. The simultaneous releases suggest coordinated availability timing, though the models serve different use cases—image generation versus text processing.

Model cards for all three releases remain sparse on technical specifications. Pricing information, detailed benchmarks, and recommended deployment configurations have not yet been published.

What This Means

The FP8 quantization of Ideogram-4 signals growing industry focus on reducing inference costs and memory footprint for large generative models. For image generation models in particular, quantization can enable deployment on single GPUs rather than requiring multi-GPU setups.

Google's Gemma-4-12B release fills a gap between smaller 2B-7B models and larger 27B+ variants, potentially offering a sweet spot for developers balancing capability and cost. The instruction-tuned version should be immediately usable for applications, while the base model allows custom fine-tuning.

Developers should wait for comprehensive benchmarks and pricing details before making production deployment decisions.