⚡ 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.
Qwen3.5-122B-A10B APEX GGUF
APEX (Adaptive Precision for EXpert Models) quantizations of Qwen3.5-122B-A10B.
Brought to you by the LocalAI team | APEX Project | Technical Report
Benchmark Results
All measurements on 8xRTX PRO 6000 Blackwell (768 GB VRAM). Perplexity on wikitext-2-raw, context 512. Accuracy benchmarks via llama.cpp (400 tasks each).
| Configuration | Size (GB) | Perplexity | KL mean | HellaSwag | Winogrande | MMLU | ARC | tg128 (t/s) |
|---|---|---|---|---|---|---|---|---|
| Q8_0 (Unsloth) | 121 | 4.819 | 0.004 | 85.5% | 77.3% | 44.19 | 57.19 | 85.5 |
| Q5_K_S (Unsloth) | ~81 | 4.826 | 0.007 | 85.3% | 76.0% | 43.80 | 57.86 | 90.4 |
| UD-Q4_K_XL (Unsloth) | ~72 | 4.829 | 0.010 | 84.8% | 76.3% | 44.25 | 55.85 | 91.8 |
| APEX I-Balanced | 83.4 | 4.831 | 0.008 | 85.5% | 77.8% | 43.86 | 57.86 | 96.7 |
| APEX I-Quality | 72.3 | 4.838 | 0.012 | 85.3% | 77.3% | 43.86 | 56.86 | 99.7 |
| APEX Quality | 72.3 | 4.848 | 0.013 | 85.5% | 76.3% | 44.44 | 55.52 | 99.8 |
| APEX Balanced | 83.4 | 4.840 | 0.008 | 85.0% | 76.3% | 43.93 | 56.86 | 96.7 |
| APEX I-Compact | 55.1 | 4.978 | 0.041 | 84.5% | 77.5% | 44.06 | 57.86 | 106.3 |
| APEX Compact | 55.1 | 5.046 | 0.049 | 84.5% | 77.8% | 43.54 | 56.19 | 106.2 |
| APEX I-Mini | 44.9 | 5.306 | 0.102 | 84.0% | 75.3% | 42.83 | 56.52 | 110.0 |
Highlights
- APEX I-Balanced matches or beats Q8_0 on HellaSwag (85.5%), Winogrande (77.8% vs 77.3%), and ARC (57.86 vs 57.19) while being 31% smaller and 13% faster.
- APEX I-Quality (72.3 GB) beats UD-Q4_K_XL at the same size on HellaSwag (85.3% vs 84.8%), Winogrande (77.3% vs 76.3%), and ARC (56.86 vs 55.85).
- APEX I-Compact (55.1 GB) achieves 84.5% HellaSwag and 57.86 ARC at 55% less size than Q8_0 — fastest standard profile at 106 t/s.
- APEX I-Mini (44.9 GB) is the smallest at 63% less size than Q8_0, still 84% HellaSwag, fastest at 110 t/s.
- I-variants consistently improve over standard profiles across PPL, KL, and ARC.
Available Files
| File | Profile | Size | Best For |
|---|---|---|---|
| Qwen3.5-122B-A10B-APEX-I-Balanced.gguf | I-Balanced | 83.4 GB | Best overall -- matches Q8_0 quality at 31% less size |
| Qwen3.5-122B-A10B-APEX-I-Quality.gguf | I-Quality | 72.3 GB | Best quality at ~72 GB tier |
| Qwen3.5-122B-A10B-APEX-Quality.gguf | Quality | 72.3 GB | Highest MMLU (44.44) |
| Qwen3.5-122B-A10B-APEX-Balanced.gguf | Balanced | 83.4 GB | General purpose, low KL |
| Qwen3.5-122B-A10B-APEX-I-Compact.gguf | I-Compact | 55.1 GB | Consumer multi-GPU, best quality/size ratio |
| Qwen3.5-122B-A10B-APEX-Compact.gguf | Compact | 55.1 GB | Consumer multi-GPU setups |
| Qwen3.5-122B-A10B-APEX-I-Mini.gguf | I-Mini | 44.9 GB | Smallest viable, fastest inference |
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: Qwen3.5-122B-A10B (Qwen3.5-MoE)
- Layers: 48
- Experts: 256 routed + 1 shared (8 active per token)
- Total Parameters: 122B
- Active Parameters: ~10B per token
- APEX Config: 5+5 symmetric edge gradient across 48 layers
Run with LocalAI
local-ai run mudler/Qwen3.5-122B-A10B-APEX-GGUF@Qwen3.5-122B-A10B-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.