Gemma-4-E4B-it — GGUF

The Gemma-4 12B model is a different architecture (gemma4_unified, 640-dim audio encoder) and is not supported by this backend.

Available variants

Model details

• Architecture: USM Conformer audio encoder (12L, 1024d, chunked-local attention with relative position bias, LightConv1d, ClippableLinear with QAT scalars) + Gemma4 LLM decoder (42L, 2560d, GQA 8Q/2KV, per-layer embeddings, hybrid sliding/full attention, GeGLU)
• Parameters: ~4B effective (~8B with embeddings)
• Audio: Gemma4AudioFeatureExtractor — 128-bin mel, 16 kHz, frame_length=320, hop=160, fft_length=512, semicausal padding, log(mel + mel_floor=0.001), no normalisation
• Languages: 140+ (ASR + speech translation)
• License: Apache 2.0
• Source: google/gemma-4-E4B-it

Usage

What's included vs an upstream Gemma-4 GGUF

• 12-layer audio conformer encoder.
• Gemma4MultimodalEmbedder audio→LLM adapter (embed_audio.embedding_projection, pre-projection RMSNorm).
• Mel filterbank + Hann window baked into the GGUF.

Provenance and EU AI Act Art. 53 note

• Upstream model: google/gemma-4-E4B-it — published by google.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.