Gemma 4 26B-A4B MoE GGUF
Certified GGUF precision quantization ladder for google/gemma-4-26b-a4b mixture-of-experts architecture.
Certified GGUF precision variants (Q8_0, Q4_K_M, Q2_K) of google/gemma-4-26b-a4b evaluated across standard reasoning, mathematics, and science benchmarks (ARC-MC, GSM8K, MATH-500) under reproducible seed controls.
Architecture
- Parameters
- 25.23B
- Active
- 3.80B
- Layers
- 30
- Quantization ladder
- Certified GGUF
- Experts
- 128
- Experts before
- —
- Top-k
- 8
Downloads
Exact release files
3 immutable artifacts
gemma-4-26b-a4b.q4_k_m.gguf
Size: 17 GB
20–24 GB reference capacity. Daily development and general conversational use.
SHA-256 Checksum
b828701dba8637250b9901a3546050ffdf3425ce3f17e6ed175a4cd158a0ba50
gemma-4-26b-a4b.q8_0.gguf
Size: 27 GB
32–48 GB reference capacity. Candidate only, not a measured fit guarantee.
SHA-256 Checksum
2c90eb5e2c0225b94602f4c76a20044c34692f998d4c75bb39a2d42bd7d244c2
gemma-4-26b-a4b.q2_k.gguf
Size: 11 GB
12–16 GB reference capacity. Memory-constrained workstations.
SHA-256 Checksum
1f2a8f06212390058f4d892d568b4331c02fd2637938e9159380d16c287b749a
Deployment
Quickstart & CLI
llama.cpp & Ollama
Stream weights directly from Hugging Face via Ollama CLI:
ollama run hf.co/Omega-Factory/gemma-4-26b-a4b-GGUF:Q4_K_MRun inference locally using llama.cpp with the recommended artifact:
llama-cli -hf Omega-Factory/gemma-4-26b-a4b-GGUF -m gemma-4-26b-a4b.q4_k_m.gguf -p "Explain the core mechanics of mixture-of-experts models." -n 512Launch a local OpenAI-compatible HTTP server:
llama-server -hf Omega-Factory/gemma-4-26b-a4b-GGUF -m gemma-4-26b-a4b.q4_k_m.gguf --port 8080 -ngl 99Evidence
Measured outcomes & precision ladder
Precision & Accuracy Matrix vs. FP16 Baseline
Holdout evaluation split (N=200 per category, seed 42) comparing all quant variants to unquantized baseline.
ARC-MC
92.0%
GSM8K
73.0%
MATH-500
36.5%
ARC-MC
92.0%
GSM8K
70.5%
-2.5 pp
MATH-500
32.5%
-4.0 pp
ARC-MC
92.5%
+0.5 pp
GSM8K
66.0%
-7.0 pp
MATH-500
23.0%
-13.5 pp
ARC-MC
82.0%
-10.0 pp
GSM8K
4.5%
-68.5 pp
MATH-500
7.5%
-29.0 pp
Item-Level McNemar Discordance & Flips vs. FP16 Baseline
Per-item analysis measuring individual test questions that flipped answer status relative to the unquantized FP16 baseline (N=200).
Recommended Release Rung Scoped Verification (Q4_K_M)
Formal cryptographic release marker evaluation records with 95% confidence intervals and sample counts.
ARC-Challenge Science Reasoning
gemma-4-26b-a4b-arc-mc-v1 · accuracy_pct
Base
92.00%
Release
92.50%
Delta
+0.50 pp
Holdout evaluation set (N=200, seed 42) comparing Q4_K_M quantized release against uncompressed FP16 baseline under standard multiple-choice log-likelihood scoring.
95% CI [-4.74, 5.74] pp · 200 samples · 1 seeds · artifact b828701dba8637250b9901a3546050ffdf3425ce3f17e6ed175a4cd158a0ba50
GSM8K Grade School Math
gemma-4-26b-a4b-gsm8k-v1 · accuracy_pct
Base
73.00%
Release
66.00%
Delta
-7.00 pp
Holdout evaluation set (N=200, seed 42) comparing Q4_K_M quantized release against uncompressed FP16 baseline under greedy decoding.
95% CI [-16.00, 2.00] pp · 200 samples · 1 seeds · artifact b828701dba8637250b9901a3546050ffdf3425ce3f17e6ed175a4cd158a0ba50
MATH-500 Advanced Mathematics
gemma-4-26b-a4b-math500-v1 · accuracy_pct
Base
36.50%
Release
23.00%
Delta
-13.50 pp
Holdout evaluation set (N=200, seed 42) comparing Q4_K_M quantized release against uncompressed FP16 baseline under greedy decoding.
95% CI [-22.36, -4.64] pp · 200 samples · 1 seeds · artifact b828701dba8637250b9901a3546050ffdf3425ce3f17e6ed175a4cd158a0ba50
Literature & Architecture
Foundational Research & Citations
Primary scientific publications & architecture preprints
Gemma 4 Technical Report
Gemma Team, Google DeepMind
Details the Gemma 4 Mixture-of-Experts architecture (26B total, 4B active parameters per token) engineered for compute-efficient edge deployment and STEM reasoning.
Gemma: Open Models Based on Gemini Research and Technology
Gemma Team, Google DeepMind
Foundational architecture, safety evaluations, and training methodology derived from Google Gemini research and technology.
Which Quantization Should I Use? A Unified Evaluation of llama.cpp Quantization
Empirical GGUF Research Study
Comprehensive empirical analysis of llama.cpp GGUF quantization schemes (Q8_0, Q4_K_M, Q2_K) across reasoning benchmarks, memory footprints, and CPU/GPU throughput.
Limitations
- RESEARCH_USE_ONLY — This model inherits a community license for research and free use; no commercial-serving claims are made.
- VRAM_ESTIMATE_NOTICE — Artifact size is disk usage, not a measured VRAM requirement. Deployment requires peak-memory profiling.
- COMPRESSED_DERIVATIVE — Pruned and re-quantized derivative of google/gemma-4-26b-a4b. Credit belongs to original authors.
Public release identity
- Release ID
- gemma-4-26b-a4b-gguf-v1
- Release version
- 1
- Published
- 2026-10-11T03:03:32.000Z
- Model ID
- Omega-Factory/gemma-4-26b-a4b-GGUF
- Base model ID
- google/gemma-4-26b-a4b
- Catalog family
- Not supplied
- Quantization rung
- Not supplied
- Supersedes
- Not supplied
- HF repository
- Omega-Factory/gemma-4-26b-a4b-GGUF
- HF repository URL
- https://huggingface.co/Omega-Factory/gemma-4-26b-a4b-GGUF
- HF revision
- e18ebc356b0534a34777fbb22a12e48aa3a62070
- HF revision URL
- https://huggingface.co/Omega-Factory/gemma-4-26b-a4b-GGUF/tree/e18ebc356b0534a34777fbb22a12e48aa3a62070
- Release manifest SHA-256
- 8ee1bc67e35fbb9a0be76599f6b9557fcff67a783688177df92a5ad6c0e86b03
- Catalog SHA-256
- e87595a7c9a5d0a6ae3662985e99356632737f2a673b773c938e4f93b214a99a
- Catalog generated
- 2026-10-11T12:45:00Z
- Commercial use
- DISALLOWED
- License URL
- https://ai.google.dev/gemma/terms