dahus/gemma-4-e2b-it-Q8_0-GGUF

Gemma 4 E2B it — Q8 GGUF

8-bit quantized GGUF version of google/gemma-4-e2b-it.
Highest quality quantization — 96% Top-1 agreement with F16, effectively lossless.

Other quantizations in this series:
Q2_K · Q3_K_S · Q3_K_M · Q4_K_S · Q4_K_M · Q5_K_S · Q5_K_M

---

File Info

Property	Value
Format	GGUF Q8
File size	4.63 GB
Bits per weight	~8
Size vs F16	1.9× 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	16.2	37.1 dB	95.7%	0.0151
🧠 Logic	16.2	36.8 dB	96.6%	0.0166
💻 Code	16.3	37.8 dB	97.4%	0.0155
🔬 Science	16.3	36.7 dB	94.2%	0.0209
Overall	16.2	37.11 dB	96.0%	0.0171

---

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
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 (this)	4.63 GB	16.2	2.9×	37.11 dB	96.0%	0.0171

---

Key Findings

• Quality: 37.11 dB SQNR and KL divergence of just 0.017 — for all practical purposes, identical to F16
• Top-1 Agreement: 96.0% — the model picks the same token as F16 96 times out of 100
• Speed: 16.2 tok/s — still 2.9× faster than F16, just slower than lower-bit quants
• Size: 4.63 GB — fits in 6 GB RAM; half the size of F16
• vs Q6_K: +8.4 dB SQNR and +1.9% Top-1 for 1.05 GB extra; worth it if you have the RAM
• Best for: Maximum quality with reasonable size reduction; production deployments where output must match F16 as closely as possible

---

Usage

# llama.cpp CLI
./llama-cli -m gemma-4-e2b-q8.gguf -p "Explain how a transformer neural network works." -n 200

# llama-cpp-python
from llama_cpp import Llama

llm = Llama(model_path="gemma-4-e2b-q8.gguf", n_ctx=2048)
output = llm("Explain how a transformer neural network works.", max_tokens=200)
print(output["choices"][0]["text"])

---

Hardware

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