gemma-4-e2b-it-Q3_K_S-GGUF

Gemma 4 E2B it — Q3_K_S GGUF

3-bit small quantized GGUF version of google/gemma-4-e2b-it.
Slightly smaller and faster than Q3_K_M with comparable output quality.

Other quantizations in this series:
Q2_K · Q3_K_M · Q4_K_S · Q4_K_M · Q5_K_S · Q5_K_M · Q6_K · Q8

---

File Info

---

Benchmark Results

Tested across 4 categories (Math, Logic, Code, Science), 3 prompts each.
Greedy decoding, 200 max new tokens. Metrics compare logit distributions vs F16 baseline.

Results by Category

---

Quantization Comparison

---

Key Findings

• Quality: 63.2% Top-1 agreement — same as Q3_K_M at the aggregate level, with lower KL divergence (1.26 vs 1.67)
• Speed: 28.9 tok/s — slightly faster than Q3_K_M (27.4 tok/s)
• Size: 2.90 GB — 80 MB smaller than Q3_K_M
• vs Q3_K_M: Q3_K_S has lower SQNR but better KL divergence, meaning the probability distributions are actually closer to F16 despite lower signal quality — practically equivalent for most tasks
• Best for: Same use case as Q3_K_M; prefer Q3_K_S if you need a tiny speed advantage

---

Usage

# llama.cpp CLI
./llama-cli -m gemma-4-e2b-q3ks.gguf -p "Explain the water cycle." -n 200

# llama-cpp-python
from llama_cpp import Llama

llm = Llama(model_path="gemma-4-e2b-q3ks.gguf", n_ctx=2048)
output = llm("Explain the water cycle.", max_tokens=200)
print(output["choices"][0]["text"])

---

Hardware

Tested on: CPU inference (llama.cpp)
Context: 2048 tokens | Greedy decoding