⚡ 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.
Carnice MoE 35B-A3B APEX GGUF
APEX (Adaptive Precision for EXpert Models) quantizations of samuelcardillo/Carnice-MoE-35B-A3B.
Brought to you by the LocalAI team | APEX Project
Available Files
| File | Profile | Size | Best For |
|---|---|---|---|
| Carnice-MoE-35B-A3B-APEX-I-Quality.gguf | I-Quality | 21 GB | Highest quality with imatrix |
| Carnice-MoE-35B-A3B-APEX-Quality.gguf | Quality | 21 GB | Highest quality standard |
| Carnice-MoE-35B-A3B-APEX-I-Balanced.gguf | I-Balanced | 24 GB | Best overall quality/size ratio |
| Carnice-MoE-35B-A3B-APEX-Balanced.gguf | Balanced | 24 GB | General purpose |
| Carnice-MoE-35B-A3B-APEX-I-Compact.gguf | I-Compact | 16 GB | Consumer GPUs, best quality/size |
| Carnice-MoE-35B-A3B-APEX-Compact.gguf | Compact | 16 GB | Consumer GPUs |
| Carnice-MoE-35B-A3B-APEX-I-Mini.gguf | I-Mini | 13 GB | Smallest viable, fastest inference |
| Carnice-MoE-35B-A3B-F16.gguf | F16 | 65 GB | Full precision reference |
Benchmark Results (Native Evals)
| Model | Size | PPL ↓ | KL ↓ | HellaSwag | WinoGrande | MMLU | ARC-C | TruthfulQA | pp512 t/s | tg128 t/s |
|---|---|---|---|---|---|---|---|---|---|---|
| F16 (ref) | 65G | 6.16 | - | - | - | - | - | - | 2315 | 109.1 |
| APEX-Quality | 21G | 6.2 | 0.010 | 83.5 | 74.0 | 40.9 | 56.9 | 34.0 | 4717 | 134.2 |
| APEX-I-Quality | 21G | 6.2 | 0.009 | 83.0 | 75.0 | 40.3 | 55.5 | 34.3 | 4734 | 132.6 |
| APEX-Balanced | 24G | 6.2 | 0.007 | 83.0 | 73.8 | 41.1 | 54.5 | 33.8 | 4572 | 130.3 |
| APEX-I-Balanced | 24G | 6.2 | 0.006 | 83.5 | 74.8 | 40.6 | 54.2 | 34.0 | 4539 | 128.7 |
| APEX-Compact | 16G | 6.4 | 0.045 | 82.8 | 75.5 | 40.8 | 55.9 | 34.0 | 4516 | 132.1 |
| APEX-I-Compact | 16G | 6.3 | 0.032 | 83.0 | 73.8 | 41.2 | 56.2 | 34.9 | 4352 | 130.6 |
| APEX-I-Mini | 13G | 6.6 | 0.071 | 82.0 | 72.2 | 40.6 | 53.8 | 33.7 | 4293 | 133.1 |
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.
Architecture
- Base Model: samuelcardillo/Carnice-MoE-35B-A3B
- Architecture: Qwen3.5-MoE 35B-A3B
- Layers: 40
- Experts: 256 routed (8 active per token)
- Total Parameters: 35B
- Active Parameters: ~3B per token
- APEX Config: 6+6 symmetric edge gradient across 40 layers
- Calibration: v1.2 diverse dataset
Run with LocalAI
local-ai run mudler/Carnice-MoE-35B-A3B-APEX-GGUF@Carnice-MoE-35B-A3B-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.