dahus/gemma-4-e2b-it-Q3_K_M-GGUF

Gemma 4 E2B it — Q3_K_M GGUF

3-bit medium quantized GGUF version of google/gemma-4-e2b-it.
Usable quality at very small size — a step up from Q2_K with only 0.2 GB more.

Other quantizations in this series:
Q2_K · Q3_K_S · Q4_K_S · Q4_K_M · Q5_K_S · Q5_K_M · Q6_K · Q8

---

File Info

Property	Value
Format	GGUF Q3_K_M
File size	2.98 GB
Bits per weight	~3
Size vs F16	2.9× 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	27.4	13.7 dB	66.1%	1.4605
🧠 Logic	27.7	14.1 dB	59.5%	2.0306
💻 Code	27.3	15.1 dB	65.7%	1.4831
🔬 Science	27.3	12.8 dB	61.3%	1.7248
Overall	27.4	13.93 dB	63.2%	1.6747

---

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	2.99 GB	31.6	5.6×	5.85 dB	32.0%	4.1149
Q3_K_S	3.11 GB	28.9	5.1×	10.12 dB	63.2%	1.2605
Q3_K_M (this)	2.98 GB	27.4	4.8×	13.93 dB	63.2%	1.6747
Q4_K_S	3.37 GB	25.0	4.4×	19.10 dB	80.9%	0.3456
Q4_K_M	3.43 GB	24.0	4.2×	20.33 dB	82.4%	0.3356
Q5_K_S	3.6 GB	21.9	3.9x	23.32 dB	87.7%	0.1547
Q5_K_M	3.63 GB	22.0	3.9×	23.25 dB	86.9%	0.1248
Q8	4.97 GB	16.2	2.9×	37.11 dB	96.0%	0.0171

---

Key Findings

• Quality: 63.2% Top-1 agreement — coherent outputs, but noticeably different from F16 on complex prompts
• Speed: 27.4 tok/s — 4.8× faster than F16
• Size: 2.98 GB — fits under 4 GB RAM with room to spare
• vs Q3_K_S: Q3_K_M has higher SQNR (13.93 vs 10.12 dB) but higher KL divergence (1.67 vs 1.26); Q3_K_S is slightly faster. In practice, quality is similar.
• Best for: Low-RAM devices where Q4 variants don't fit; simple chat and factual Q&A

---

Usage

# llama.cpp CLI
./llama-cli -m gemma-4-e2b-q3km.gguf -p "Explain the water cycle." -n 200

# llama-cpp-python
from llama_cpp import Llama

llm = Llama(model_path="gemma-4-e2b-q3km.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