⚡ Each donation = another big MoE quantized
I host 25+ free APEX MoE quantizations as independent research. My only local hardware is an NVIDIA DGX Spark (122 GB unified memory), enough for ~30-50B-class MoEs, but bigger ones (200B+) require rented compute on H100/H200/Blackwell, typically $20-100 per quant.
If APEX quants are useful to you, your support directly funds those bigger runs.
🎉 Patreon (Monthly) | ☕ Buy Me a Coffee | ⭐ GitHub Sponsors
💚 Big thanks to Hugging Face for generously donating additional storage, much appreciated.
Gemma 4 26B-A4B APEX GGUF
APEX (Adaptive Precision for EXpert Models) quantizations of google/gemma-4-26B-A4B-it.
Brought to you by the LocalAI team | APEX Project | Technical Report
Benchmark Results
Benchmarks coming soon (re-quantized with llama.cpp b8664 including Gemma 4 tokenizer and logit softcapping fixes). For reference APEX benchmarks on the Qwen3.5-35B-A3B architecture, see mudler/Qwen3.5-35B-A3B-APEX-GGUF.
Available Files
| File | Profile | Size | Best For |
|---|---|---|---|
| gemma-4-26B-A4B-APEX-I-Balanced.gguf | I-Balanced | 19 GB | Best overall quality/size ratio |
| gemma-4-26B-A4B-APEX-I-Quality.gguf | I-Quality | 20 GB | Highest quality with imatrix |
| gemma-4-26B-A4B-APEX-Quality.gguf | Quality | 20 GB | Highest quality standard |
| gemma-4-26B-A4B-APEX-Balanced.gguf | Balanced | 19 GB | General purpose |
| gemma-4-26B-A4B-APEX-I-Compact.gguf | I-Compact | 15 GB | Consumer GPUs, best quality/size |
| gemma-4-26B-A4B-APEX-Compact.gguf | Compact | 15 GB | Consumer GPUs |
| gemma-4-26B-A4B-APEX-I-Mini.gguf | I-Mini | 13 GB | Smallest viable, fastest inference |
| mmproj.gguf | Vision projector | 1.2 GB | Required for image understanding |
What is APEX?
APEX is a quantization strategy for Mixture-of-Experts (MoE) models. It classifies tensors by role (routed expert, shared expert, attention) and applies a layer-wise precision gradient -- edge layers get higher precision, middle layers get more aggressive compression. I-variants use diverse imatrix calibration (chat, code, reasoning, tool-calling, agentic traces, Wikipedia).
See the APEX project for full details, technical report, and scripts.
Architecture
- Model: Gemma 4 26B-A4B (google/gemma-4-26B-A4B-it)
- Layers: 30
- Experts: 128 routed (8 active per token)
- Total Parameters: 26B
- Active Parameters: ~4B per token
- Vision: Built-in vision encoder (mmproj included)
- APEX Config: 5+5 symmetric edge gradient across 30 layers
- Calibration: v1.3 diverse dataset (chat, code, reasoning, multilingual, tool-calling, Wikipedia)
- llama.cpp: Built with b8664 (includes Gemma 4 tokenizer fix, logit softcapping, newline split)
Run with LocalAI
local-ai run mudler/gemma-4-26B-A4B-it-APEX-GGUF@gemma-4-26B-A4B-APEX-I-Balanced.gguf
Credits
APEX is brought to you by the LocalAI team. Developed through human-driven, AI-assisted research. Built on llama.cpp.