Skip to main content
Back

Google's Gemma 4 E4B Now Runs on Any Laptop After 994K Downloads

A community GGUF conversion of Google's multimodal Gemma 4 E4B-it lands on Hugging Face, unlocking local llama.cpp inference across laptops and consumer GPUs.

Pro

Model, Full Access
  • ggml-org published an automated GGUF conversion of Google's Gemma 4 E4B-it on Hugging Face.
  • Available in Q4_0 (4.59 GB), Q8_0 (8.03 GB) and BF16 (15.1 GB) with matching MTP heads.
  • Runs out of the box in llama.cpp, Ollama, LM Studio, Jan and Docker Model Runner.
  • Base model is 8B params, Apache 2.0, with 128K context and multimodal text/image/audio input.
  • Downloads have crossed 994k, making it a popular local-first Gemma 4 build.
  • Best fit for on-device chat, OCR, voice prototypes and function-calling agents.

An automated GGUF conversion of Google's Gemma 4 E4B-it is now live on Hugging Face under the ggml-org organization, giving llama.cpp users a drop-in, quantized build of one of the smaller members of Google's newest open model family. The upload has crossed 994k downloads, making it one of the more actively pulled Gemma 4 variants in the community, even though Google DeepMind published the underlying weights earlier.

The upload repackages Google's instruction-tuned E4B checkpoint into the GGUF container format used by llama.cpp, Ollama, LM Studio, and Jan, produced automatically by the ggml-org/convert pipeline. That distinction matters because the repackaging is what makes the model runnable on a laptop CPU or a single consumer GPU without touching the original PyTorch weights.

What actually shipped

The repo publishes three precision tiers of the same E4B-it checkpoint, along with matching Multi-Token Prediction (MTP) heads. Size determines whether it fits in your VRAM:

QuantizationModel sizeMTP head
Q4_0 (4-bit)4.59 GB59.7 MB
Q8_0 (8-bit)8.03 GB98.7 MB
BF16 (16-bit)15.1 GB172 MB

The reported parameter count is 8B under the gemma4 architecture identifier, and the license passes through as Apache 2.0. Source checkpoints include the standard