dahus/gemma-4-e2b-it-Q4_K_M-GGUF

Gemma 4 E2B it — Q4_K_M GGUF

4-bit medium quantized GGUF version of google/gemma-4-e2b-it.
Recommended default — best balance of quality, speed, and size in the series.

Other quantizations in this series:
Q2_K · Q3_K_S · Q3_K_M · Q4_K_S · Q5_K_S · Q5_K_M · Q6_K · Q8

---

File Info

Property	Value
Format	GGUF Q4_K_M
File size	3.43 GB
Bits per weight	~4
Size vs F16	2.7× smaller

---

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

Category	Speed (tok/s)	SQNR	Top-1 Agreement	KL Divergence
🔢 Math	24.3	19.9 dB	84.5%	0.2786
🧠 Logic	23.9	20.3 dB	81.5%	0.4190
💻 Code	23.3	20.6 dB	80.5%	0.3618
🔬 Science	24.7	20.6 dB	83.3%	0.2830
Overall	24.0	20.33 dB	82.4%	0.3356

---

Quantization Comparison

Model	Size	Speed (tok/s)	vs F16 speed	SQNR	Top-1 Agree	KL Div
F16 (baseline)	8.67 GB	5.7	1.0×	baseline	baseline	baseline
Q2_K	2.78 GB	31.6	5.6×	5.85 dB	32.0%	4.1149
Q3_K_M	2.98 GB	27.4	4.8×	13.93 dB	63.2%	1.6747
Q4_K_S	3.13 GB	25.0	4.4×	19.10 dB	80.9%	0.3456
Q4_K_M (this)	3.43 GB	24.0	4.2×	20.33 dB	82.4%	0.3356
Q5_K_S	3.35 GB	21.9	3.9×	23.32 dB	87.7%	0.1547
Q5_K_M	3.38 GB	22.0	3.9×	23.25 dB	86.9%	0.1248
Q6_K	3.58 GB	19.9	3.5×	28.72 dB	94.1%	0.0743
Q8	4.63 GB	16.2	2.9×	37.11 dB	96.0%	0.0171

---

Key Findings

• Quality: 82.4% Top-1 agreement — large jump from Q3 (63%), outputs are reliable and coherent
• Speed: 24.0 tok/s — 4.2× faster than F16
• Size: 3.19 GB — fits in 4 GB RAM
• vs Q4_K_S: Q4_K_M is marginally better quality (20.33 vs 19.10 dB SQNR, 82.4 vs 80.9% Top-1) for only 60 MB more
• Best for: General-purpose use — math, code, Q&A, chat; the go-to choice when you don't have a specific reason to pick another variant

---

Usage

# llama.cpp CLI
./llama-cli -m gemma-4-e2b-q4km.gguf -p "Write a Python function for binary search." -n 200

# llama-cpp-python
from llama_cpp import Llama

llm = Llama(model_path="gemma-4-e2b-q4km.gguf", n_ctx=2048)
output = llm("Write a Python function for binary search.", max_tokens=200)
print(output["choices"][0]["text"])

---

Hardware

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