# APEX I-Quant vs Unsloth UD vs Standard K-Quant — Quality Comparison

Date: 2026-07-03
Models tested: Qwen3.5-35B-A3B (architecture same as Qwen3.6-35B-A3B)
Hardware: RTX 2080 Ti 11 GB (sm75)

## Key Finding

**APEX I-Compact is NOT better quality than standard Q4_K_M.** The APEX advantage is smaller file size and MTP baked in, not quality.

## Source Data

### From APEX Technical Report (Qwen3.5-35B-A3B on DGX Spark)

| Configuration | Size | PPL | MMLU | KL mean | Speed (tg128) |
|---|---|---|---|---|---|
| F16 (reference) | 64.6 GB | 6.537 | 41.5% | — | 30.4 t/s |
| Q8_0 | 34.4 GB | 6.533 | 41.2% | 0.0046 | 52.5 t/s |
| Unsloth UD-Q4_K_L | 18.8 GB | 6.586 | 41.1% | 0.0151 | 65.5 t/s |
| bartowski Q4_K_M | 19.77 GB | **6.610** | ~41% | 0.0182 | — |
| bartowski Q3_K_M | 15.1 GB | 6.730 | 41.5% | 0.0420 | 60.6 t/s |
| **APEX Compact (non-I)** | 16.1 GB | **6.783** | 40.9% | 0.0469 | 69.8 t/s |
| **APEX I-Compact** | 16.1 GB | **6.669** | 41.7% | 0.0332 | 69.8 t/s |

### From Unsloth Qwen3.5-35B-A3B GGUF Benchmarks

| Quant | Size | PPL | KL 99.9% | KL mean |
|---|---|---|---|---|
| bartowski Q4_K_M | 19.77 GB | 6.6097 | 0.5771 | 0.0182 |
| bartowski IQ4_XS | 17.42 GB | 6.6234 | 0.7265 | 0.0234 |
| noctrex MXFP4_MOE | 20.55 GB | 6.5948 | 0.7939 | 0.0248 |
| Unsloth UD-Q4_K_XL | 19.17 GB | **6.5918** | **0.4097** | **0.0137** |
| Unsloth UD-Q5_K_XL | 23.22 GB | 6.5489 | 0.2360 | 0.0069 |

## Rankings by Quality

**Best → Worst (by perplexity):**

1. **Unsloth UD-Q5_K_XL** (23.2 GB) — PPL 6.549
2. **Unsloth UD-Q4_K_XL** (19.2 GB) — PPL 6.592
3. **Unsloth UD-Q4_K_L** (18.8 GB) — PPL 6.586
4. **noctrex MXFP4_MOE** (20.6 GB) — PPL 6.595
5. **bartowski Q4_K_M** (19.8 GB) — PPL 6.610
6. **bartowski IQ4_XS** (17.4 GB) — PPL 6.623
7. **APEX I-Compact** (16.1 GB) — PPL 6.669
8. **bartowski Q3_K_M** (15.1 GB) — PPL 6.730
9. **APEX Compact (non-I)** (16.1 GB) — PPL 6.783

## Recommendations for 11 GB VRAM

### Target: Maximum Quality
- **Unsloth UD-Q4_K_XL** (19.2 GB) — best quality that fits, but tight on 11 GB with context
- **Unsloth UD-IQ4_NL** (~18.5 GB) — Dynamic 2.0 "no-loss" IQ4, better than Q4_K_M, smaller

### Target: Best Quality/Size Tradeoff
- **APEX I-Compact** (17.3 GB with MTP) — ≈Q4_K_M quality, MTP baked in, 6x faster cold load
- **Unsloth UD-IQ4_NL** (18.5 GB) — slightly better quality than Q4_K_M, MTP baked in

### Target: Smallest Plus MTP
- **APEX I-Mini** (14.3 GB) — quality between Q3 and Q4, but with MTP
- **APEX I-Nano** (11.7 GB) — aggressive, comparable to Q2-Q3 quality, but with MTP

## Real-World Benchmark on RTX 2080 Ti

| Model | Size | Gen (warm) | VRAM | Load |
|---|---|---|---|---|
| Qwen3.6 Q4_K_M (Unsloth) | 22 GB | 56.5 t/s | 10,874 MiB | 61s |
| Qwen3.6 APEX I-Compact | 17.3 GB | 51.8 t/s | 9,710 MiB | 10s |

The I-quant compute penalty on sm75 costs ~5 t/s (10%) compared to K-quants at similar quality. The tradeoff is faster load and 1.2 GB more VRAM headroom for context.