# Fleet Benchmarks — June 20, 2026 (after --n-cpu-moe optimization)

All benchmarks on RTX 2080 Ti (11 GB) using llama.cpp b9743 (sm75-optimized build).

## Optimized Fleet (6 Models)

| # | Model | t/s | Context | VRAM | Quant | File Size | --n-cpu-moe | Gain |
|---|-------|-----|---------|------|-------|-----------|-------------|------|
| 1 | **gemma-12b** | 96.2 | 131K | 9.0 GB | Q4_K_XL | ~7 GB | N/A (dense) | — |
| 2 | **gemma-26b-200k** | 63.5 | 200K | 10.1 GB | Q4_K_XL | ~15 GB | 16 | — |
| 3 | **qwen36-35b-mtp** | 60.0 | 250K | 9.2 GB | Q4_K_M | 22 GB | 32 | +17% |
| 4 | **nemotron-term-14b** | 56.2 | 32K | 10.6 GB | Q4_K_S | ~8 GB | N/A (dense) | — |
| 5 | **GLM-4.7-flash-reap** | 53.5 | 128K | 9.1 GB | Q4_K_M | 14 GB | 20 | +72% |
| 6 | **qwen3-coder-next-80b** | 35.7 | 128K | 8.4 GB | Q4_K_M | 45 GB | 42 | +8% |

## Removed Models

| Model | Reason | File Freed |
|-------|--------|------------|
| deepseek-v2-lite | Redundant (coder-30b covers same speed with 4x context) | 12 GB |
| gemma4-v2-agentic | Covered by coder-30b + qwen36 | 7 GB |
| qwen3-coder-30b | Redundant with coder-next-80b + qwen36 | 18 GB |
| qwen3-30b-thinking-eagle3 | Redundant with qwen36 (slower + same Eagle3 limited by CPU offload) | 18 GB |
| nemotron-3-nano | Redundant with GLM (similar speed, less context) | 23 GB |
| qwen3-coder-next-reap | REAP-pruned, slower than coder-next-80b | 32 GB |
| qwen3.5-122b-a10b | Too slow (12.5 t/s), 55 GB wasted | 55 GB |
| Moonlight-16B | Native 8K context too limited | 12 GB |
| Devstral-24B | Too slow (6.6 t/s) | 15.6 GB |
| DiffusionGemma | Needs special build, Q4_K_M too large | 16 GB |

Total disk freed: ~208 GB