gemma-4-e2b-Q6_K-GGUF

Gemma 4 E2B — Q6_K GGUF

6-bit quantized GGUF version of google/gemma-4-e2b-it.
Near-lossless quantization — 94% Top-1 agreement with F16 at a fraction of the size.

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 · 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: 94.1% Top-1 agreement — crosses the "near-identical to F16" threshold; only Q8 is better
• SQNR: 28.72 dB — a substantial 5.5 dB jump over Q5_K_M; outputs are essentially indistinguishable from F16 in practice
• Speed: 19.9 tok/s — 3.5× faster than F16
• Size: 3.58 GB — only 1 GB more than Q4_K_M for a major quality improvement
• vs Q8: Q6_K is 1.05 GB smaller and 3.7 tok/s faster, with only a small quality difference (94.1 vs 96.0% Top-1)
• Best for: Quality-sensitive tasks where you want near-F16 fidelity but still need to fit under ~5 GB; scientific explanation, precise reasoning, complex multi-step code

---

Usage

# llama.cpp CLI
./llama-cli -m gemma-4-e2b-q6k.gguf -p "Explain the difference between supervised and reinforcement learning." -n 200

# llama-cpp-python
from llama_cpp import Llama

llm = Llama(model_path="gemma-4-e2b-q6k.gguf", n_ctx=2048)
output = llm("Explain the difference between supervised and reinforcement learning.", max_tokens=200)
print(output["choices"][0]["text"])

---

Hardware

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