# Fleet Benchmarks — June 20, 2026 (Optimized)

Binary: llama-server-sm75, b9743 (c57607016), CUDA 12.6, sm75 Turing
GPU: RTX 2080 Ti 11GB | CPU: Xeon E5-2697A v4 16C/32T

All models optimized with `--n-cpu-moe` tuning for ~9 GB VRAM target.

| Profile | Model | t/s | VRAM | Ctx | Cold Start | n-cpu-moe |
|---------|-------|-----|------|-----|-----------|-----------|
| gemma-12b | Gemma 4 12B QAT Q4_K_XL | **96.2** | 9.0 GB | 131K | 17s | n/a (dense) |
| gemma-26b-200k | Gemma 4 26B A4B QAT Q4_K_XL | **63.5** | 10.1 GB | 200K | 37s | 16 |
| qwen36-35b-mtp | Qwen3.6-35B-A3B-MTP Q4_K_M | **60.0** | 9.2 GB | 250K | 19s | **32** |
| nemotron-term-14b | Nemotron-Terminal-14B Q4_K_S | **56.2** | 10.6 GB | 32K | 5s | n/a (dense) |
| glm-4.7-flash-reap | GLM 4.7 Flash REAP 23B Q4_K_M | **53.5** | 9.1 GB | 128K | 13s | **20** |
| nemotron-3-nano | Nemotron-3-Nano-30B Q4_K_M | **46.8** | ~9 GB | 128K | 35s | **38** |
| qwen3-coder-30b | Qwen3-Coder-30B-A3B Q4_K_M | **38.3** | 10.1 GB | 128K | 37s | **42** |
| qwen3-coder-next | Qwen3-Coder-Next-80B Q4_K_M | **35.7** | 8.4 GB | 128K | 133s | **42** |
| qwen3-30b-thinking-eagle3 | Qwen3-30B-A3B-Thinking + Eagle3 Q4_K_M | **35.9** | 10.3 GB | 128K | 45s | 128 |

## Removed

| Model | Reason |
|-------|--------|
| deepseek-v2-lite | Redundant (coder-30b covers same speed, better context) |
| gemma4-v2-agentic | Role covered by coder-30b (38 t/s) and qwen36 (60 t/s) |
| REAP-48B | Slower than coder-30b (29 vs 38 t/s) |
| Qwen3.5-122B-A10B | Too slow (12.5 t/s) for daily use |