Gemma 4 E4B

Name: Gemma 4 E4B
Author: Google

Google · Gemma 4 · released 2026-04-02 · data checked 2026-11-01

Gemma 4 E4B is a 8B-parameter dense model from Google (United States), released 2026-04-02 under the Apache-2.0 license. It scores 38.2 on the Small LM Index, #20 of 26 small models. It needs about 5.6 GB of memory at 4-bit with an 8K-token context.

4.5B effective parameters (8B including per-layer embeddings). Built for on-device use.

Small LM Index

38.2

#20 of 26 scored small models

Likely 34.1–42.4 · from 4 of 6 core benchmarks

Parameters

4.5B effective

Price per 1M tokens

—

No hosted API on OpenRouter; self-host

Context window

128K

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best small model on that benchmark. Sources: 6 reported by Google. How we label sources.

MMLU-Pro Lab-reported Knowledge69.4

Class best: Qwen3.5 9B 82.5

GPQA Diamond Lab-reported Knowledge58.6

Class best: Qwen3.5 9B 81.7

AIME Lab-reported Math · AIME 2026, no tools42.5

Class best: Apriel 1.6 15B Thinker 88.0

LiveCodeBench Lab-reported Coding · LiveCodeBench v652.0

Class best: Granite 4.2 8B 73.2

τ²-Bench Lab-reported Agentic42.2

Class best: Qwen3.5 9B 79.1

MMMU-Pro Lab-reported Vision52.6

Class best: Qwen3.5 9B 69.2

Not reported: estimated

What the index model expects from this model's other scores. Not a measurement.

Humanity's Last Exam≈ 4

SWE-bench Verified≈ 18

Specifications

Lab: Google (US · United States)
Architecture: Dense transformer
Total parameters: 8B
Active parameters: All (dense)
Size class: Small
Context window: 128K tokens
Input modalities: text, image, audio
Reasoning: Hybrid (thinking on/off)
License: Apache-2.0
Openness: Open weights
Memory to run: ~5.6 GB at 4-bit, 8K context (more sizes)
Hugging Face: 4.1M downloads (30 days) · 1.8K likes

Weights on Hugging Face ↗

Compare with

Gemma 4 E4B vs Granite 4.2 8B (8B, index 63.5)
Gemma 4 E4B vs Ministral 3 8B (8B, index 55.2)
Gemma 4 E4B vs DeepSeek-R1-0528-Qwen3-8B (8.2B, index 53.1)
Gemma 4 E4B vs Qwen3 8B (8.2B, index 48.5)
Gemma 4 E4B vs Llama 3.1 8B (8B, index 21.3)
Gemma 4 E4B vs Ling 3.0 Tiny (7.9B, index 51.9)

Sources

https://huggingface.co/google/gemma-4-31B-it

Something wrong here? Report it.

Run it locally

Uses ggml-org/gemma-4-E4B-it-GGUF (Q4_0). Swap the tag for another size: BF16, Q4_0, Q8_0.

Ollama

ollama run hf.co/ggml-org/gemma-4-E4B-it-GGUF:Q4_0

llama.cpp

llama-server -hf ggml-org/gemma-4-E4B-it-GGUF:Q4_0

Then open http://localhost:8080 for a chat window.

MLX (Mac)

pip install -U mlx-lm
mlx_lm.chat --model mlx-community/gemma-4-e4b-it-4bit

Text only; for images use mlx-vlm.

vLLM (GPU server)

vllm serve google/gemma-4-E4B-it

Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization	4K context	32K context	128K context
16-bit F16	17 GB	17 GB	18 GB
8-bit Q8_0	9.2 GB	9.5 GB	10 GB
4-bit Q4_K_M	5.6 GB	5.8 GB	6.6 GB

Check it against your GPU, Mac or phone →