GE

Gemma 4 12B It Heretic NVFP4

GemmaGGUF by MagnarksUpdated 2026-10-13
Download on Hugging Face
Parameters11.9B
Context256K
Layers48
Architecturegemma4
Smallest quant7.0 GB
Quick answer

Gemma 4 12B It Heretic NVFP4 needs about 20.4 GB of memory at its main file (a 7.0 GB file) with an 8K context window, so it fits on a 24 GB GPU. This repository has 2 quants, from 7.0 GB to 8.2 GB. Longer context needs more: at 32K tokens the KV cache alone adds about 51.5 GB.

Gemma 4 12B It Heretic NVFP4 – a quantized multimodal LLM

Gemma 4 12B It Heretic NVFP4 is a 11.9‑billion parameter model optimized for text, image, and audio input. It uses FP4 weight and activation quantization with GPTQ calibration, reducing disk size and GPU memory by about 65%. The model is released under the apache‑2.0 license.

Good at

  • Multimodal input handling
  • FP4 quantization for lower VRAM usage
  • Stable inference with vLLM

Limits

  • No vision support in llama.cpp by default
  • Requires NVFP4 quantized weights
  • Large VRAM needed for 8k context

Best for

  • Running on NVIDIA Blackwell GPUs
  • Speech models alongside multimodal LLM
  • Low‑latency inference with reduced disk footprint

Who it’s for. Developers needing a multimodal LLM that fits on RTX 50 series GPUs. The model is suitable for applications that combine text, image, and audio processing while keeping VRAM usage manageable.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
—7.0 GB 20.4 GB
Too big Get file →
—8.2 GB 21.5 GB
Too big Get file →

For 8 GB VRAM, the 6.98 GB file is the smallest option, estimating 20.4 GB of VRAM at 8k context – still above 8 GB. For higher VRAM budgets, the 8.16 GB file uses about 21.5 GB at 8k context. Both estimates are approximate and depend on batch size and other factors.

What GPU do you need for Gemma 4 12B It Heretic NVFP4?

These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.

  • 4 GB VRAM: no quant fits at an 8K context.
  • 6 GB VRAM: no quant fits at an 8K context.
  • 8 GB VRAM: no quant fits at an 8K context.
  • 12 GB VRAM: no quant fits at an 8K context.
  • 16 GB VRAM: no quant fits at an 8K context.
  • 24 GB VRAM: the best quant that fits is the main file (8.2 GB file, about 21.5 GB in use).

Specifications

Architecture from the GGUF header

Architecturegemma4
Parameters11.9B
Layers48
Trained context262,144 tokens
KV cache (fp16)1536 KB per token

Source

Quantized byMagnarks
FamilyGemma
Licenseapache-2.0
Downloads (HF)1,699
Last updated2026-10-13

Frequently asked questions about Gemma 4 12B It Heretic NVFP4

How much VRAM does Gemma 4 12B It Heretic NVFP4 need?

About 20.4 GB at its main file with an 8K context window. That is the 7.0 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.

Can I run Gemma 4 12B It Heretic NVFP4 on an 8 GB GPU?

Not at an 8K context. The smallest file needs about 20.4 GB, so you would need a larger GPU or CPU offloading.

What is the file size of Gemma 4 12B It Heretic NVFP4 GGUF?

It depends on the quant: file is 7.0 GB, file is 8.2 GB.

What context length does Gemma 4 12B It Heretic NVFP4 support?

The GGUF header lists a trained context length of 262,144 tokens (256K). The KV cache grows with the context you use: about 12.9 GB at 8K and 51.5 GB at 32K.

Where can I download Gemma 4 12B It Heretic NVFP4 GGUF?

The files are on Hugging Face in the Magnarks/gemma-4-12B-it-heretic-NVFP4-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.

What license does Gemma 4 12B It Heretic NVFP4 use?

Hugging Face lists the license as apache-2.0. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.

Can I run this model on 8 GB VRAM?

No, the model requires roughly 20–21 GB of VRAM at 8k context, which exceeds 8 GB.

Does this model support vision in llama.cpp?

The GGUF conversion keeps vision projections in BF16, but llama.cpp does not enable vision by default.

What license applies to this model?

The model is released under the apache‑2.0 license.

Similar models

Browse all →

More from Magnarks

All by Magnarks →

More Gemma models

All Gemma →

Other medium models

All medium models →