dahus/gemma-4-e2b-it-Q2_K-GGUF

Gemma 4 E2B it — Q2_K GGUF

2-bit quantized GGUF version of google/gemma-4-e2b-it.
Smallest and fastest variant in the series — use only if RAM is the hard constraint.

Other quantizations in this series:
Q3_K_S · Q3_K_M · Q4_K_S · Q4_K_M · Q5_K_S · Q5_K_M · Q6_K · Q8

---

File Info

Property	Value
Format	GGUF Q2_K
File size	2.78 GB
Bits per weight	~2
Size vs F16	3.1× 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	30.9	5.0 dB	35.3%	3.8922
🧠 Logic	31.7	5.7 dB	33.8%	4.1991
💻 Code	31.8	6.7 dB	24.4%	4.4969
🔬 Science	32.0	6.4 dB	34.3%	3.8713
Overall	31.6	5.85 dB	32.0%	4.1149

---

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 (this)	2.78 GB	31.6	5.6×	5.85 dB	32.0%	4.1149
Q3_K_S	2.90 GB	28.9	5.1×	10.12 dB	63.2%	1.2605
Q3_K_M	2.98 GB	27.4	4.8×	13.93 dB	63.2%	1.6747
Q4_K_M	3.19 GB	24.0	4.2×	20.33 dB	82.4%	0.3356
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: Significant degradation — only 32% Top-1 agreement with F16; output can be incoherent (see sample below)
• Speed: 31.6 tok/s — fastest in the series, 5.6× faster than F16
• Size: 2.78 GB — fits in under 4 GB RAM
• Best for: Extreme RAM-constrained environments where some output quality loss is acceptable; not recommended for reasoning or code tasks

⚠️ Warning: Q2_K produces visibly broken outputs on this model. Sample response to a math prompt repeated token garbage (skills skills skills...). Consider Q3_K_S or higher for usable results.

---

Usage

# llama.cpp CLI
./llama-cli -m gemma-4-e2b-q2k.gguf -p "Solve step by step: 2 + 2 = ?" -n 200

# llama-cpp-python
from llama_cpp import Llama

llm = Llama(model_path="gemma-4-e2b-q2k.gguf", n_ctx=2048)
output = llm("Solve step by step: 2 + 2 = ?", max_tokens=200)
print(output["choices"][0]["text"])

---

Hardware

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