---
---

# Fleet Overhaul — July 3, 2026 (Final)

## Summary

Complete fleet overhaul: added repeat penalties to all models, replaced Ornith-9B dense with MTP variant, redownloaded Qwythos v3 (looping fix), replaced Ornith-35B Q6 with APEX-I-Compact, removed GLM-4.7-Flash. Tried Qwen3.6 APEX I-Compact and UD-IQ4_NL — both slower than original Q4_K_M on sm75. Reverted Qwen to original Q4_K_M. Upgraded Ornith-9B from Q5→Q6.

## Changes Applied

### 1. Repetition Fix (all models)

**Problem:** No `--repeat-penalty` on any model. Default 1.0 = disabled. Small dense 9B models loop without it. Ornith-9B-Q5 was repeating itself and misidentifying as Qwen 35B MoE.

**Root cause:** GGUF metadata has `general.architecture = qwen35`. The model sees "qwen35" and hallucinates being the 35B MoE variant.

**Fix:**
- Dense (≤9B): `--repeat-penalty 1.10 --repeat-last-n 256`
- MoE (≥26B): `--repeat-penalty 1.05 --repeat-last-n 256`
- All: `--min-p 0.05 → 0.10`

### 2. Ornith-9B: Dense → MTP-Q5 → MTP-Q6

Dense 69.7 t/s → MTP Q5 71.3 t/s → MTP Q6 69.1 t/s. Q6 slightly slower on sm75 due to memory bandwidth. Kept Q6 for quality.

### 3. Qwythos v3

Redownloaded v3 from empero-ai. Fixed looping during long generation. File size identical (7.6 GB).

### 4. Ornith-35B Q6 → APEX-I-Compact

28 GB Q6 (23.8 t/s, 64s) → 17 GB APEX (28.5 t/s, 12s). 5.3x load, +15% gen, 2x context.

### 5. GLM-4.7-Flash Removed

18 GB, 29.7 t/s, no MTP.

### 6. Qwen3.6: Three Options Tested

| Quant | Size | Warm t/s | Result |
|---|---|---|---|
| Q4_K_M (original) | 22 GB | 56.5 | **Kept** — fastest |
| APEX I-Compact MTP | 17.3 GB | 51.8 | 9% slower, 1.2 GB VRAM saved |
| UD-IQ4_NL | 18.5 GB | 38.7 | 32% slower — reverted |

**Key finding:** On sm75, K-quants > I-quants > UD Dynamic 2.0 for speed.

## Final Fleet (5 models)

| Model | Quant | Size | Ctx | Gen2 | VRAM |
|---|---|---|---|---|---|
| qwythos-9b-mtp-q6 | Q6_K MTP | 7.1 GB | 131K | 96.4 | 10,458 |
| ornith-9b-mtp-q5 | Q6_K MTP | 7.6 GB | 96K | 69.1 | 9,650 |
| qwen36-35b-mtp | Q4_K_M MTP | 22 GB | 230K | 56.5 | 10,874 |
| gemma-26b-200k | Q4_K_XL + MTP | 14 GB | 230K | 55.8 | 10,502 |
| ornith-35b-q6-mtp | APEX-I-Compact MTP | 17 GB | 262K | 28.5 | 9,900 |

Flags across all: `--repeat-penalty 1.05/1.10`, `--repeat-last-n 256`, `--min-p 0.10`, `--spec-draft-n-max 2-3`.