Gemma-4-E4B-text-base

Google Gemma-4-E4B-text-base (Modality-Stripped Text Backbone)

Gemma-4-E4B-text-base is a clean, text-only extraction of Google's Gemma-4-E4B-it instruction-tuned model. By decoupling and stripping multimodal projections (vision and audio towers), this checkpoint provides a pure language model backbone optimized for text, software engineering, and tool-calling deployments.

---

⚡ Why Modality-Stripped?

Standard Gemma-4 checkpoints integrate multimodal encoders and cross-attention projectors. For dedicated text generation, reasoning, and coding agent workloads:

1. VRAM Reduction: Saves memory otherwise reserved for vision transformer buffers and projector parameters.
2. Simplified Inference: Eliminates multimodal input pre-processing overhead and dependencies in downstream serving engines (vLLM, TensorRT-LLM, llama.cpp).
3. Pristine Weight Fidelity: The underlying text transformer parameters (model.language_model.*) are 100% identical bit-for-bit to the upstream release.

---

🔬 Model Profile

• Parameter Count: $\sim 4.2\text{B}$ parameters.
• Context Length: 131,072 tokens.
• Format: Standard Hugging Face safetensors compatible with transformers >= 4.40.0.
• Vocabulary Size: 256,000 tokens.

---

⚠️ Important: Tool-Calling Parser Bugfix in vLLM

• Upstream Bug Report: vllm-project/vllm#54256
• Official Pull Request: vllm-project/vllm#54257 — [Bugfix][Parser] Support bare call: and whitespace-free channel transitions in Gemma4 parser

---

🚀 Serving with vLLM

vllm serve wyattearp/Gemma-4-E4B-text-base \
  --host 0.0.0.0 \
  --port 8000 \
  --dtype bfloat16 \
  --max-model-len 131072 \
  --tool-call-parser gemma4 \
  --reasoning-parser gemma4 \
  --enable-auto-tool-choice

---

📄 License

Inherits upstream Gemma Terms of Use from Google DeepMind.