# Fleet Model Specs — June 27, 2026

RTX 2080 Ti 11GB, 62GB system RAM, Xeon E5-2697A v4. All models via llama-swap :9292.

## Full Fleet (9 models)

| Model ID | File | Size | `--n-cpu-moe` | TG | Notes |
|---|---|---|---|---|---|
| `qwen36-35b-mtp` | Qwen3.6-35B-A3B-MTP-Q4_K_M | 22 GB | 32 | 76 t/s | MTP 97.5% accept, maxed VRAM |
| `gemma-26b-200k` | gemma-4-26B-A4B-it-qat-UD-Q4_K_XL + MTP draft | 14+0.24 GB | 16 | 86 t/s | 256K ctx, +MTP draft |
| `nemotron-term-14b` | Nemotron-Terminal-14B.Q4_K_S | 8 GB | none (dense) | 56 t/s | Fastest prompt ingestion |
| `glm-4.7-flash` | GLM-4.7-Flash-Q4_K_M | 18 GB | 24 | 36 t/s | 200K ctx, `--reasoning off` required |
| `gpt-oss-20b` | gpt-oss-20b-Q6_K | 12 GB | none (MoE falls through) | 61 t/s | Lightest MoE variant |
| `ornith-35b-q6-mtp` | ornith-1.0-35b-Q6_K-MTP-final (skinnyctax Frankenstein) | 28 GB | 30 | ~20-25 t/s | MTP grafted, main `ornith` alias |
| `ornith-35b-q8` | ornith-1.0-35b-Q8_0 | 35 GB | 32 | 18 t/s | Fallback, alias `ornith-q8` |
| `qwen-agentworld-35b-q6` | Qwen-AgentWorld-35B-A3B-UD-Q6_K | 28 GB | 28 | 20.5 t/s | World model, no spec |
| `ornith-9b-q6` | ornith-1.0-9b-Q6_K | 6.9 GB | none (dense) | 76 t/s | Full GPU, dense 9B |

## All Models in llama-swap Config

```yaml
  qwen36-35b-mtp:              aliases: [qwen36, qwen36-mtp, primary]
  gemma-26b-200k:              aliases: [gemma, gemma-26b, gemma-26b-200k, gemma-200k, deep-reason, high-ctx]
  nemotron-term-14b:           aliases: [nemotron-14b, term-14b, terminal-14b]
  glm-4.7-flash:               aliases: [glm-4.7, glm-flash, glm]
  gpt-oss-20b:                 aliases: [gpt-oss, oss-20b]
  ornith-35b-q6-mtp:           aliases: [ornith, ornith-35b, ornith-mtp]
  ornith-35b-q8:               aliases: [ornith-q8]
  qwen-agentworld-35b-q6:      aliases: [agentworld, agentworld-35b, aw35b]
  ornith-9b-q6:                aliases: [ornith-9b]
```

## n-cpu-moe Benchmarks (RTX 2080 Ti 11GB)

### Ornith-1.0-35B-Q8_0 (35 GB)
| n-cpu-moe | VRAM | TG | OOM? |
|---|---|---|---|
| 40 | 2.6 GB | 17.0 t/s | No |
| 38 | 4.2 GB | 17.5 t/s | No |
| 36 | 5.8 GB | 16.2 t/s | No |
| 32 | 9.1 GB | 18.0 t/s | No ← best |
| 28 | — | — | OOM |

### Ornith-1.0-35B-Q6_K-MTP (28 GB)
| n-cpu-moe | Result |
|---|---|
| 30 | Stable (in production) |
| lower | Not benchmarked in this session |

### Qwen-AgentWorld-35B-A3B-UD-Q6_K (28 GB)
| n-cpu-moe | VRAM | TG | OOM? |
|---|---|---|---|
| 32 | ~9 GB | 19.9 t/s | No |
| 28 | — | 20.5 t/s | No ← best |
| 24 | — | — | OOM |
| 20 | — | — | OOM |
| 16 | — | — | OOM |

### Pattern: Q6 models (28 GB) need n-cpu-moe ≥ 28; Q8 (35 GB) needs ≥ 32