# Fleet Benchmarks — 2026-06-20

## Binary: llama.cpp master b9743 (c57607016), sm75-optimized

| # | Model | t/s | VRAM | Ctx | Cold Start | Arch | Notes |
|---|-------|-----|------|-----|-----------|------|-------|
| 1 | gemma-12b | 96.2 | 9.0 GB | 131K | 17s | gemma3 | Upstream regression (#24514), was 118 |
| 2 | gemma4-v2-agentic | 52.7 | 7.7 GB | 131K | 3s | gemma4_unified | Best at 16K (53), use 131K for longer work |
| 3 | gemma-26b-200k | 63.5 | 10.1 GB | 200K | 37s | gemma3 MoE | Deep reasoning |
| 4 | qwen36-35b-mtp | 51.4 | 6.8 GB | 250K | 19s | qwen35moe | Best coder, draft-mtp active |
| 5 | nemotron-term-14b | 56.2 | 10.6 GB | 32K | 5s | qwen3 dense | Terminal specialist |
| 6 | nemotron-3-nano-30b | 36.4 | 3.4 GB | 128K | 35s | nemotron_h MoE | VRAM efficiency king |
| 7 | qwen3-coder-30b | 34.6 | 7.9 GB | 128K | 37s | qwen3moe | No spec decoding |
| 8 | qwen3-coder-next-80b | 33.1 | 4.4 GB | 128K | 2:13 | qwen3next hybrid | 2 min cold start |
| 9 | qwen3-coder-reap-48b | 29.4 | 6.6 GB | 128K | 1:06 | qwen3next pruned | REAP 40% pruned |
| 10 | GLM-4.7-Flash-REAP-23B | 31.1 | 2.6 GB | 32K | 13s | GLM arch | Runs alongside others |
| 11 | Qwen3-30B-A3B-Thinking | 34.8 | 7.9 GB | 128K | 45s | qwen3moe | Pairs with Eagle3 |

## New Models Not Yet Benchmarked
- Qwen3.5-122B-A10B (UD-Q3_K_M, multi-part) — download in progress
- Eagle3 speculator (safetensors) — needs GGUF conversion + pairing

## Removed Models
- DeepSeek-Coder-V2-Lite — redundant with coder-30b, removed 2026-06-20